VOLUME 92 NUMBER 1 JANUARY 1985
Levels of Processing, Encoding Specificity, Elaboration, and CHARM
Janet Metcalfe Eich University of British Columbia, Vancouver, British Columbia, Canada
A model of cued recall called CHARM (composite holographic associative recall model) is applied to several findings that have been investigated within the depth- of-processing framework. It is shown that, given some straightforward, empirically testable assumptions about the representations of the to-be-remembered items themselves, CHARM can account for the main effect of depth of processing, the problem of the negatives, encoding-specificity interactions, and effects—both facilitative and inhibitory—of elaboration. The CHARM model is extended to encompass some depth-of-processing effects found in recognition memory.
One of the most interesting characteristics same in both cases, and so seems to be of human cognition is that when people attributable to the way a person processes process information, that processing is such and combines information. Although a large that the information may be changed or number of experiments have shown changes transformed. For instance, the verbal context due to context and it is generally agreed that in which a particular item is given is highly such biasing effects are fundamental to human influential in determining the interpretation cognition, our understanding of the processes of that item. If a word cat is presented in the underlying such interactive effects is far from context of lions, the feline qualities of the cat complete. are emphasized; whereas, if the same word The most interesting property of the mem- cat is presented in the context of dogs, the ory model called CHARM (composite holo- domesticated animal characteristics of the cat graphic associative recall model; Eich, 1982; are liable to be more salient. This result Metcalfe & Murdock, 1981; and see also occurs even though the stimulus may be the Borsellino & Poggio, 1973; Cavanagh, 1976; Gabor, 1968; Julesz & Pennington, 1965; Liepa, 1977; Longuet-Higgins, 1968; Mur- This research was facilitated by a Natural Sciences dock, 1982, 1983; van Heerden, 1963; Will- and Engineering Research Council (NSERC) postdoctoral fellowship, NSERC Grant A0505, and a University of shaw, 1981, for related models) is that the California, Los Angeles, intramural computing grant to operations for association formation, storage, the author. and retrieval in this model cause changes in I thank James Anderson, Robert Bjork, Patrick Cavan- the representation of items from input to agh, Fergus Craik, Eric Eich, Ronald Fisher, Thomas memory to output from memory. These Nelson, Anne Treisman, Endel Tulving, Thomas Wickens and four anonymous reviewers for their critiques and changes are sometimes just degradations of assistance. I thank Morris Moscovitch for providing data the initial input. Under some more interesting and materials relevant to Simulation 1, and Ronald conditions, though, the model causes system- Fisher for providing the materials as well as the design atic biasing. This biasing may affect the in- for Experiment 1. Requests for reprints should be sent to Janet Metcalfe terpretation of items as well as the level of Eich, Department of Psychology, University of Brit- recall. ish Columbia, Vancouver, British Columbia, Canada In the CHARM model, items are character- V6T 1Y7. ized as patterns of features. These patterns
Copyright 1985 by the American Psychological Association, Inc. 0033-295X/85/S00.75
1 JANET METCALFE EICH may vary in their similarity to one another. of experiments that show the memorial con- The representations of items are associated sequences of varying the context in which by means of the operation of convolution items are presented have been conducted and stored in a composite memory trace in within the levels-of-processing framework which other associations are superimposed (Cermak & Craik, 1979; Craik & Lockhart, as well. In order to retrieve, the representation 1972). Within this framework, the context in of a cue item is correlated with the composite which an item is presented is often manipu- trace, resulting in a noisy and sometimes lated by giving subjects different orienting distorted item. This retrieved item may be tasks. For instance, a subject might be asked identified as a word—if the task is verbal— whether a target word rhymed with another by being matched to every item in a semantic word, or alternatively, whether a word was in lexicon. Figure 1 gives a schematic illustration a particular category. Such differences in of the CHARM model. The operations in the context result in pronounced effects on the boxes in the figure (i.e., convolution, addition, level of recall of targets. Another program of and correlation) are responsible for changes research (Thomson & Tulving, 1970; Tulving in the nature of the items that depend on & Osier, 1968; Tulving & Thomson, 1973) context. These operations are discussed in showed that it is not only the context at time more detail shortly. of study that is important, but also the context In this article I explore some of the me- at time of test. For instance, if a target item morial consequences of the changes in rep- CAT was encoded in the context of lions, but resentation—depending on context—that are is later probed with the cue dog, recall is predicted by the CHARM model. A number much poorer than if the initial context item
FROM PERCEPTUAL ANALYSIS
TO PATTERN IDENTIFICATION
COMPOSITE MEMORY TRACE (ADD)
Figure 1. A diagrammatic overview of CHARM. LEVELS AND CHARM is used. The nature of the elaborating context (shallowly). Many experiments have con- in which an item is presented also influences firmed that the manipulation of an orienting the recall of the item. Sometimes elaboration task produces a main effect on the goodness results in enhanced recall, and sometimes of recall or recognition. Of particular interest elaboration results in impaired recall. I pro- in this article are experiments that have pose that the memory processes in the CHARM produced not only a main effect of orienting model and the resultant changes in represen- task (i.e., level of processing), but also quali- tation are sufficient to account for many of fying interactions. The qualifying interactions the experimental results of levels of process- have given rise to a number of constructs ing, encoding specificity, and elaboration. such as elaboration (J. R. Anderson & Reder, Although I propose that many of the levels- 1979; Battig, 1979; Craik & Tulving, 1975; of-processing, test context, and elaboration Moscovitch & Craik, 1976; Ross, 1981), cue- effects are produced by the associative, stor- trace compatibility (Tulving, 1979; and see age, and retrieval mechanisms in CHARM, the also Bransford, Franks, Morris, & Stein, model was not originally designed with these 1979), and distinctiveness (Craik, 1979; Craik effects in mind. Rather the model was in- & Jacoby, 1979; Eysenck, 1979; Jacoby & tended to address the general question of how Craik, 1979; Klein & Saltz, 1976; Lockhart, it is that people associate ideas, store those Craik, & Jacoby, 1976; Nelson, 1979) that associations in memory, and then later, when have been evoked instead of or in addition given one idea, are able to retrieve another to the original depth hypothesis. However, as from memory. I have applied the model to a Jacoby and Dallas (1981) note, "Although wide range of human learning, memory, and there are currently a large number of exper- classification phenomena in the cued recall, iments showing effects of manipulating an or paired-associate paradigm (see Eich, 1982), orienting task, there is no generally accepted as well as to rehearsal and free recall of framework that incorporates the results of unrelated items (Metcalfe & Murdock, 1981). those experiments" (p. 309). In free recall, many effects appear to depend In none of the situations that are modeled on factors other than the mechanisms of using computer simulations of CHARM, do I association formation, storage, and retrieval. assume a priori that the initial encoding of For example, the availability of retrieval cues an item differs from one condition to another. due to the subject's state, and to primary To be more specific, I do not assume that (short-term) memory were important in the perceptual analysis produces different patterns holographic free-recall model, in a manner of features to represent the same word, CAT, similar to other free-recall models (J. R. for instance, before association formation Anderson, 1972; Atkinson & Shiffrin, 1968; and retrieval, depending on the experimental Gillund & Shiffrin, 1984; Raiijmakers & Shif- condition. Treisman (1979) reviewed the per- frin, 1981). Although of considerable interest ceptual literature that indicates that the re- and importance in their own right, these lation between perceptual operations and factors are nevertheless extraneous to the memory performance may not be as straight- transformational properties of composite hol- forward as was originally supposed. Jacoby ographic models that are of main interest and Dallas (1981) argued that if the levels here, so I mainly consider cued recall. Later manipulation has its effect by causing different in the article, I also sketch out how the main perceptual analyses—for instance, that a effect of depth of processing could be obtained shallow task results in the encoding of only in recognition, within the CHARM model. a few letters or phonemes, whereas a deep The original levels-of-processing hypothesis task results in the encoding of the entire stated that the memorability of a particular word—an effect of level of processing should target item depends on the depth of perceptual be found in a perceptual identification task analysis performed on that item (Craik & in which the subject must name a tachisto- Lockhart, 1972). Items that were analyzed to scopically presented word. In contrast to this a semantic level (deeply) were assumed to be prediction, they showed that perceptual iden- more memorable than were items that were tification of words that had been presented analyzed phonemically or orthographically in a levels task was not differentially affected JANET METCALFE EICH by the levels manipulation. A sizable levels unrelated items. Items that are similar to one effect on memory performance, however, was another are represented such that the feature found in the same experiment. Presumably, values are not independent; that is, there is all of the items, regardless of the levels ma- some feature overlap, beyond independence, nipulation, are perceptually analyzed as among similar items. Figure 2 gives an illus- meaningful, and the memory effects have tration of feature profiles of four items, A, some other cause. B, C, and D. In the figure, Items A and B Because of the interactive properties of the are unrelated; the values on each feature were CHARM model, differences in the retrieved randomly drawn from a normal distribution patterns of features, that is, in the character- centered on zero. I have introduced some istics of the output from episodic memory, similarity between Items B and C. As can be are produced depending on the characteristics seen from the figure, Features -10 to —1 of the items that are associated and used as have identical values in Items B and C. Items retrieval cues. These differences are produced C and D are also similar to one another by means of the operations that are proposed because Features 0 to 10 have identical values. to be the operations that people use to asso- Similarity need not be coded as identical ciate ideas, store those associations in mem- values but, because it will simplify the ex- ory, and then later to retrieve them. position, I use this coding in the matrices that follow. Description of CHARM The dot product, which is a measure of the similarity between two items, A and B, • In this section, a brief encapsulation and is given by some illustrative examples of CHARM are (B-l)/2 given. This model has been described previ- A - B = 2 a,A, (D ously (Eich, 1982). Items in the model are /=-(«-0/2 represented as ordered sets of features (Tver- sky, 1977; D. Wickens, 1972). The features where Item A is coded as are coded as numerical values that may be either positive or negative, but within a par- ticular item are assumed to have values that -i, a0,al,a2,..., a d0, di, d2, •) and D = (• • •, d-2, d-\, two items by means of the operation of do, dt, d2, convolution is conceptually distinct from the similarity of the items. As an example, two Convolution objects (e.g., telephones or computers) may When people associate two items (e.g., the be quite similar to one another without in- words ape and book) it is assumed that the teracting or being connected or associated in ordered sets of features representing those any way. The result of association by convo- items are combined by the operation of con- lution is an interactive new entity (vector) volution. Convolution, denoted *, is denned that does not, in any obvious way, correspond as to the items that were so combined. (A*B)m = 2 a,bj, (2) Figure 3 gives the vector that results from (IJ)fS(m) the convolution of Items A and B in Figure n- 1 2. The dimensions in the vector resulting where S(m) = {(i, j) from convolution do not correspond to the features in the item vectors. For instance, n-\ , and / + j = m\. The association of suppose that the central (og) feature in the items represents some quality such as size ... -10 -9 -8 -7 -6 -5 -4 -3 -2 -1 0 1 2 3 4 5 6 7 8 9 10 ... F E A T U R E S Figure 2. Feature profiles of four items that vary in their similarity to one another. JANET METCALFE EICH A * B Figure 3. The feature profile that results from the convolution of Items A and B in Figure 2. (with the numerical values being the sizes of The components of the convolution of the the particular items that are represented). two 5-feature items are shown in Figure 4. The central feature of the convolution vector Substituting into the matrix, it can be seen is not size, but rather is a combination of all that the dimensions (£_4, £-3, . . .) in the of the features in the items that were con- convolution vector are not the same as are volved. the dimensions in the original items. For Figure 4 gives an illustration, in matrix instance, the central dimension in the con- form, of the convolution of two items, each volution vector is Bill's value on attractive- coded on five features. The result (t-4, t-3, ness X Bob's value on aggressiveness + Bill's t-2, t-i, t0, ti, t2, h, t4) is a single vector of value on first phoneme X Bob's value on dimensionality, 2n - 1. For the sake of illus- age + Bill's value on hair color X Bob's value tration, suppose that the items were two on hair color + Bill's value on age X Bob's words, Bill and Bob, each coded on features value on first phoneme + Bill's value on (aggressiveness, first phoneme of name, hair attractiveness X Bob's value on aggressiveness. color, age, attractiveness). Both of the items The association formed by convolution is would have a value (not necessarily the same) complex—an interaction of the original items. on each of these features (and on others If we consider the patterns of features at because it is assumed that there are a large the level of the items to be meaningful (and number of features representing each item). to be consciously interpretable), the pattern Figure 4. A matrix illustrating the convolution of two items, A = (fl-2, a_i, OQ, a\, flj), and B = (6_2, b-.t, bo, bt, hi). LEVELS AND CHARM that results from convolution is not mean- Correlation ingful to the person and is stored in memory. When a retrieval cue is given, it is assumed The problem then, of course, becomes one that the representation of the cue item is of deconvolving the memory trace so that correlated with the composite trace in order something meaningful that can be consciously to retrieve an item. The item retrieved, how- interpreted is recovered. To convert the non- ever, is not identical to any item that was meaningful and nonconscious memory trace initially encoded. The correlation, denoted #, into a form of representation that is mean- ingful, retrieval occurs. The retrieval opera- of two vectors is tion, that deconvolves the trace, is discussed = 2 x,yj, (4) shortly. The form of the trace, as something not directly available to consciousness, con- forms closely to what we generally mean by where memory—that which is not conscious. The view of the person that is implied is quite an S(m) active one because something (convolution) must be done to associate items, and a dif- ferent operation (correlation) must be per- , and i -j = formed to convert the nonmeaningful mem- ory trace into a remembrance that is mean- Figure 5 illustrates the correlation of Item ingful and consciously interpretable. B with the convolution of two unrelated items, A and B. I have left the joint convo- Addition lution-correlation matrix in an expanded form to illustrate the components that reduce The results of convolution are added into to noise (with expected values of zero) and a composite trace. There is but one associative those that produce signal; the signal compo- trace, which is the sum of the various results nents have been circled. It can be seen that of association formation. The composite trace the central n features of the resulting vector, is analogous to a photograph that has been R, produce the Item A plus noise. For in- exposed many times, except that each "ex- stance, posure" is itself a complex combination (via convolution) of two items. The trace may be denoted T, where + (noise) T = A*B + C*D (3) = a_2(B-B) + noise and A, B, C, and D are list items that have = a_z(l) + noise. been convolved. The idea of a composite trace in which representations are superim- In a similar way, r_i is a_, + noise, r0 is OQ + posed is used in a number of models (J. A. noise, and so on for all of the retrieved Anderson, 1973, 1977; J. A. Anderson, Sil- features. The retrieved item differs from the verstein, Ritz, & Jones, 1977; Cavanagh, 1976; initial Item A insofar as the retrieved item is Kohonen, Oja, & Lehtio, 1981; Metcalfe & noisy. Murdock, 1981; Murdock, 1982, 1983) and Figure 6 shows the convolution-correlation has precidents in the ideas of Gallon (1879). matrix when A is used as a cue. The result Note that under certain circumstances the is B + noise. To simplify the matrix, I have results produced by a composite trace are entered 0 for all of the components that, indistinguishable from those of multiple trace because of the assumption of statistical in- models such as those of Hintzman (1983), dependence among features, have an expected where the results of multiple retrievals are value of zero (even though these values will added. The composite trace conserves storage not be precisely zero, but will be noise). space and constrains the retrieval mechanism. Comparison of Figures 5 and 6 shows that The search metaphor, or course, cannot apply the signal components are different, depending with a composite trace. on whether A or B is correlated with the JANET METCALFE EICH LEVELS AND CHARM trace. If an item unrelated to either A or B Figure 7 gives an analogy to the result is correlated with the trace, all components produced from the addition of unrelated in the joint convolution-correlation matrix associations in the composite trace. The anal- have expected values of zero and the result ogy is not perfect because in the figure the R is only noise. dots are coded as zeros (white) and ones (black) rather than as having values centered Pattern Identification around zero. Thus, in the figure, but not the model, the addition of noise makes the re- If the task is verbal, the pattern that is sulting representation darker. Nevertheless, it retrieved by correlating a cue with the com- can be seen from the figure that the addition posite trace must be identified so that a of noise decreases the discriminability of the discrete verbal response may occur. In verbal signal. Even if only one association were recall, the retrieved item is matched against entered in the composite trace, the result of every item in the lexicon (except the probe retrieval differs from the initial item insofar or cue itself) and the best match, above a as the retrieved item is noisy. The addition minimum criterion, is the verbal item re- of noise is the simplest representational called. change that results from the holographic as- Later in the article, in the section on sociation and composite trace in CHARM. Recognition and CHARM, I propose that rec- Associative interference—the A-B A-C ognition retrieval is similar to recall, except trade-off. There are no constraints on what that the item retrieved by the probe is items can be associated with one another in matched only to the probe itself. The match- the model. The left panel of Figure 8 provides ing process involves the computation of the an analogy to the retrieved item when two dot product between the retrieved item and unrelated items are associated (A*B) and the lexical response possibilities (in recall) or then A is correlated with the trace. The the probe (in recognition). successive panels show the result of retrieval in the following situations: Some Examples A#[(A*B) + (A*C)] = B + C (+ noise) Unrelated associations add noise. Figures 5 and 6 show one association in the composite A#[A*B) + (A*C) + (A*C)] trace. Suppose, however, that the trace were = B + C + C (+ noise) constructed as A#[A*B) + (A*C) + (A*C) + (A*C)] T = A*B + C*D + E*F + = B + C + C + C(+ noise), and all of the items were unrelated. Correlat- ing A with T gives when the items A, B, and C are unrelated. It can be seen from the figure that the increasing R = A#T number of A-C associations added into the = A#[(A*B) + (C»D) + (E*F) + • • •] composite trace, causes the Item C to become more emphasized in the composite represen- = A#(A*B) +A#(C*D) tation that is retrieved by correlating A with + A#(E*F)+ • • • the trace. This situation, corresponding to the A-B, A-C paradigm (Barnes & Under- = B + noiseA.B + noisec.D wood, 1959), has been discussed in more detail in Eich (1982). + noise .F + • • •. E Prototype extraction. If items that are Each unrelated association contributes noise similar to one another are each convolved to the item that is retrieved. with the same unrelated Item, X, and each Figure 5. The matrix illustrating the correlation of Item B = (b-2, b-t, b0, bt, b^), with the trace formed by convolution as was shown in Figure 4. (The signal components that are responsible for the retrieval of A have been circled.) 10 JANET METCALFE EICH LEVELS AND CHARM 11 N C R E A S I N G NOISE Figure 7. A pictorial analogy to the addition of noise that results when unrelated associations are added into the composite trace. association is added into the composite trace, We may summarize the result of retrieval the retrieved item will still be composite in Ras form. In this situation, however, the retrieved representation will highlight the features that R = X#(T) = X#[(A*B) + (C*D) + • • • )] were common among the items. The top row + + noise . + of Figure 9 gives four items that are similar A B to one another. The bottom row gives an + noisec.D + (5) approximation to the item that is retrieved where T is the trace, A, B, C, D, and so via correlation as follows from left to right: forth, are list items, SXA is the similarity between X and A, and X is any cue item. X#(X*A) = A, Equation 5 falls directly and immediately out of the method of association formation, stor- X#[(X*A) + (X*B)] = A + B, age, and retrieval. It is important for under- X#[(X*A) + (X*B) + (X*C)] = A + B + C, standing the results that follow. Intrapair similarity. Suppose, now, that X#[(X*A) + (X*B) + (X*Q + (X*D)] the two items that are associated by convo- = A + B + C + D. lution are similar to one another. Figure 10 gives an example of the convolution matrix It can be seen that the idiosyncratic aspects produced by two similar items, of the separate items are lost and a prototype A = (a_ , a_, , OQ, a, , a ), and emerges (see J. A. Anderson, 1977; Eich, 2 2 1982; Hintzman, 1983, for further discussion). B = (a_2, fl_i,oo, &,,&2). Figure 6. The matrix illustrating the correlation of Item A = (a_2, a_i, Oo, a,, a2), with the trace formed by convolution as was shown in Figure 4. (The signal components that are responsible for the retrieval of B have been circled. Components that are only noise with an expected value of zero are indicated as zero in this illustration, in order to simplify the matrix.) 12 JANET METCALFE EICH B + C B + C + C B+C + C+C Figure 8. A pictorial analogy to the simultaneous retrieval of more than one unrelated item as in the A-B, A-C paradigm. Figure 11 shows the joint convolution-cor- that produce signal are circled, just as in relation matrix that results when B is corre- Figures 5 and 6. By comparing Figure 11 to lated with the association. The components Figure 5, it can be seen that all of the ^ ^^IWS^A: <•- J^Jlp^^feV.K-- •:^r.r.^^fe^v:.: 3sX&-£wSB$& S:#"-v • W?v A + B A-I-B + C Figure 9. A pictorial analogy to the simultaneous retrieval of highly similar items, which is responsible for prototype abstraction. LEVELS AND CHARM 13 Figure 10. A matrix illustrating the convolution of two similar items, A = (fl_2, a~\, a NOISE \V^'-:-/C;. . • •• • " • • . »!•' ...... • " ". • '•". •' . ""-X' • A +• NOISE A + B + NOISE B -I- NOISE Figure 12. A pictorial analogy to the composite retrieved when two similar items have been associated, is shown in the bottom center panel. (The left and right bottom panels give an analogy to retrieval when the items associated were not similar to one another.) graphs of two people—Millicent Gallon on model when Iwo similar items are convolved the top left, and Charles Darwin on the top and then one of them is used as a retrieval right. Darwin and Gallon were cousins and cue. For instance, have a certain family resemblance in face shape, eye spacing, and so on. The top center Darwin#(Darwin * Gallon) panel is noise. The bottom left panel shows = Gallon + SOG Darwin + noise. the result of adding noise to the depiction of Gallon alone. Such a result occurs, in the II can be seen lhal Ihe figure in Ihe bottom model, under conditions in which the two center panel stands out better, against the items convolved are unrelated, and then one noise, than does Ihe figure in eilher of Ihe item is used to cue the other, bottom side panels. X#(X* Gallon) = Gallon + noise. The presenl explanation of Ihe main effecl of levels of processing is based on Ihe testable The bottom center panel is Ihe addition of assumption lhat often the representations of Darwin, Gallon, and noise. This is an ap- Ihe deeply encoded contexl and target items proximation lo Ihe item relrieved by Ihe (e.g., Figure 11. The matrix illustrating the correlation of Item B = (a_2, a-i, a<,, i, 62) with the trace formed by convolution as in Figure 10. (The signal components have been circled.) 16 JANET METCALFE EICH another than are the representations of the Many researchers investigating the represen- shallowly encoded context and target items tation of natural and constructed categories (e.g., hat-CAT). This assumption is discussed have concluded that the similarity among in more detail in the next section. Because exemplars is of critical importance. For in- of the associative and retrieval mechanism in stance, Rosen (1977) discussed categorical the CHARM model—which produce Equation structure in terms of the real-world correla- 5—this difference in similarity, in this situa- tional structure of concrete objects. According tion, results in better recall in the deeply than to Rosch (1977) there is a natural basic level in the shallowly encoded conditions. of classification that is the most inclusive Before proceeding to some evidence about level at which lower level objects have (a) a the nature of the interitem similarities in number of attributes in common, (b) motor situations like levels experiments, one further movements in common, (c) objective similar- comment is necessary. The model does not ity in shape, and (d) a class that is identifiable predict that similarity, under any circum- from the averaged shapes of the lower level stances whatever, increases recall. To give a objects. Note that all four of these aspects few examples, if the similarity of all of the reflect what is here called similarity among items in one list of paired associates is greater the items. For each of the four aspects of the than the similarity of all of the items in a correlational structure, she has found evidence second list, the model predicts recall to be that category members (subordinates) do in- higher in the unrelated than in the highly deed overlap with one another and with the similar list. This prediction, as well as the more abstract basic level object. corresponding experimental result, was shown Sokal (1977), in discussing classification in Simulation 1 and Experiment 1 of Eich from a taxonomic point of view, noted that (1982). If all of the cue items are similar to the process of classification is critically de- one another, recall is predicted to be rather pendent on the perception of similarities. He bad because all of the items have a tendency noted that the distinction may be drawn to be retrieved, and the response selection is between monothetic classifications (those with liable to be quite inaccurate. In a like manner, defining features) and polythetic classifica- certain combinations in the Osgood (1949) tions. The latter are those in which "individ- surface with high similarity result in poor uals or objects share a large proportion of recall. The entire Osgood surface is produced their properties but do not necessarily agree by the mechanisms in the CHARM model. It in any one property . . . A corollary of is not possible to reduce the point or hypoth- polythetic classification is the requirement esis of the present article to a statement to that many properties (characteristics) be used the effect that similarity improves recall. The to classify objects" (p. 190). Sokal goes on to more accurate statement is that similarity argue, after Wittgenstein (1958), that most has an effect on the result of the transfor- natural classifications are of the polythetic mation that occurs by means of storage by type. The important point, for the present convolution and addition, and retrieval by article, is that classification of an item as a correlation. The particular similarity relation member of a particular category is likely to that is usually used in levels tasks corresponds be based on shared properties (though not to a situation in the model in which an necessarily defining shared properties). increase in similarity helps recall. The importance of similarity in classifica- tion is also evident in data on the acquisition Similarity, Levels of Processing, and CHARM of word meanings among children. The over- A major assumption in applying the extensions that children use in applying par- CHARM model to levels of processing main ticular terms strongly suggest that similarity effects is that the similarity between the con- among referents, whether it be perceptual text and the target item tends to be greater similarity (Clark, 1973), functional similarity in the deep than in the shallow conditions. A (K. Nelson, 1974), or some combination of task that is frequently given to induce deep different sorts of similarities (Bowerman, processing, is to ask subjects whether a par- 1977), is critical in classification judgments ticular item is a member of a certain category. even at a young age. LEVELS AND CHARM 17 The impact of similarity on classification central ideas was independent of whether the judgments is not restricted to natural classi- question should receive a yes or no response. fications but is also exemplified by studies They found, in this case, that there were no that have used artificial or constructed stimuli. memorial differences depending on the type Posner and Keele (1970) found that items of response. that were large distortions of the learned Semantic properties or features need not schema or exemplars were classified less well be represented differently from sensory prop- than items that were small distortions. Items erties. For instance, in Rosch's (1977) ex- that were highly similar to the constructed ample, cited above, category members were category (i.e., small distortions) tended to be shown to be similar in shape. Shape, however, classified as category members; items that would not obviously be considered to be a were less similar (i.e., large distortions) tended semantic property, although it may be used to a greater extent to be rejected as category to make a semantic judgment. As Lockhart members. It may be inferred that with in- (1979) put it, creasing dissimilarity between the target item The distinction between sensory and semantic codes has and the representation of the category, there in an important sense been drawn too sharply. There is would be an increasing tendency to say that a tendency to regard the sensory component of a word an item was not a member of that particular or picture as a kind of detachable skin within which is a category. kernel of meaning but that is itself meaningless. Sensory and semantic features do not possess this simple additive There appears to be rather good agreement relationship. The sensory codes are aspects of analysis of that there is high similarity among category meaning, (p. 82) members. Most classification models are based on this tenet in one form or another Sensory features are here considered to be (e.g., Hintzman & Ludlum, 1980; Medin & part of the representation of items, along Schaffer, 1978; Reed, 1973; Rosch & Mervis, with what might be called semantic features. 1975; and see Smith & Medin, 1981). There D. Nelson, Fosselman, and Peebles (1971) is also high similarity between the category have shown that memory performance was name and its own exemplars (Tversky & facilitated when cue-target pairs shared letters Smith, in Smith & Medin, 1981, p. 118). It in common (e.g., cactus-carrot, instep-in- therefore seems reasonable to suppose that flux). This study indicated that sensory, and when a subject is given a question in a levels- not just semantic features, have memorial of-processing task such as "Is it a fish?— implications. D. Nelson, Wheeler, and Brooks TROUT," there is considerable feature overlap (1976) combined both semantic and sensory or similarity between the mental representa- similarity in the same experiment and found tions of the key ideas, fish and TROUT. How- that both had an effect on performance. It is ever, when the person is given a categorical assumed that even with shallow processing question to which the answer is no, such as tasks, there is some similarity between the "Is it a musical instrument?—TROUT," there context and target, so long as the answer to is rather low similarity between the key ideas. the shallow question is yes. In short, if category inclusion is based on high similarity, category exclusion is based Simulation 1 on insufficient similarity. The similarity between context and target The first simulation is addressed to the appears to have memorial implications. For cued recall, rhyme, and category conditions instance, Schulman (1974) found that con- of Moscovitch and Craik's (1976) Experiment gruous questions (e.g., Is a soprano a singer?) 1. Subjects were given questions that required produce better memory performance than do either a positive or a negative response. Ex- incongruous questions (e.g., Is mustard con- amples of the four conditions were (a) cate- cave?). Craik and Tulving (1975) tested the gory positive (Is it in the category fruit? idea that it is the congruity that is important CHERRY), (b) rhyme positive (Does it rhyme and not whether a question receives a positive with pond? WAND), (c) category negative (Is or negative response. Subjects were asked it in the category office supplies? LAMP), and questions in which the congruity between the (d) rhyme negative (Does it rhyme with stock? 18 JANET METCALFE EICH CLIPS). At time of test, the subjects were Table 1 given the encoding questions and were asked Percentage Recalled as a Function of Encoding for recall of the appropriate (encoded) target Question and Type of Response words. Encoding condition The results of this experiment are presented in the top panel of Table 1. As can be seen Type of Rhyme Category from the table, there was a main effect of response (%) (%) level of processing. Cued recall was better Experiment 1* when category questions were asked than Positive 35 78 when rhyme questions were given. In addition Negative 11 18 to this main effect, there was a main effect b Simulation l of type of response (yes or no) such that yes Positive 33 80 targets were recalled better than no targets. Negative0 13 15 Qualifying both of these effects was an inter- action between level of processing and type • From "Depth of Processing, Retrieval Cues, and Unique- ness of Encoding as Factors in Recall" by M. Moscovitch of response. The effect of level of processing and F. I. M. Craik, 1976, Journal of Verbal Learning and was much attenuated with the questions that Verbal Behavior, 15, p. 449. Copyright 1976 by Academic received a no response. The difference be- Press, Inc. Adapted by permission. tween the rhyme-positive and category-posi- b Criterion = 1.0. c The difference in the negatives in the simulation is not tive conditions was 42%, whereas the differ- systematic. ence for the negative conditions was only 7%. The pattern of recall in the positive con- ditions, of course, corresponds well to the the other list items; in the rhyme-negative original depth of processing formulation. condition (stock-cups), the context and target However, as Moscovitch and Craik noted, the are unrelated to one another and to the other pattern of results for the negative responses list items. From the evidence outlined, this poses some problems. Because it is the case ordering of similarities seems to be a reason- that category questions are assumed to pro- able guess.1 The context-target pairs are as- voke deep processing and rhyme questions sociated by convolution, stored in a composite shallow processing, regardless of the answers memory trace, and retrieved by correlation. to these questions, there should have been no The retrieved item is then identified by being difference in the recallability of the targets matched to every item in the lexicon, except depending on the type of response. To address the cue item itself. The lexical item with this problem with the negatives, the ideas of which the retrieved item most strongly reso- elaboration (Craik & Tulving, 1975), unique- nates, above a certain minimal criterion, is ness or distinctiveness (Moscovitch & Craik, considered to be the response. In this simu- 1976), and congruency (Craik & Tulving, lation, the criterion for responding is varied. 1975; Moscovitch & Craik, 1976) have been Method. A lexicon of 18 items was con- evoked in addition to depth of processing as structed. All 18 items were assigned numerical being determinants of recall. In contrast, in values on each feature by randomly selecting this simulation these results are attributed to each feature value from a unit normal distri- the similarity between the context and target bution. There were 63 features in each item. items and the mechanisms in CHARM. It is assumed in this simulation that in the category-positive condition (/rw/r-CHERRY), 1 Although this guess about similarity seems reasonable it may be incorrect. In particular, the semantic-negative the context and target are highly similar to context and target items may be more similar than the one another, and unrelated to the other list rhyme negatives. Among the semantic negatives were items; in the rhyme-positive condition (pond- questions such as "round object—lamp, cherry, rock, WAND), the context and target are somewhat pool, pail, glass," which intuitively seem to exhibit some similar to one another and unrelated to the similarity. The data in several other experiments also suggest a difference in similarity between the two negative other list items; in the category-negative con- conditions. The argument proposed in Simulation 1 is dition (office supplies-LAMP), the context and not circular because the similarity values are potentially target are unrelated to one another and to available empirically, as is exemplified by Experiment 1. LEVELS AND CHARM 19 The feature values were normalized for every the means and standard deviations of the item (F) so that F • F = 1. Items 1 and 2 were resonance values were calculated, for each designated as the context and target of the lexical item. category-positive condition. To represent the Results. The frequency of correct recall similarity between them, 54 features were for the high criterion (1.0) simulation are randomly selected without replacement and presented in the bottom panel of Table 1; the were reassigned the same value in Item 2 as relevant data from Moscovitch and Craik's those features had in Item 1. Items 3 and 4 (1976) experiment are shown in the top panel. were designated as the context and target in As can be seen from the table, the simulation the rhyme-positive condition. Twenty-seven produced the main effect of levels—category features were randomly selected without re- positive over rhyme positive. It also produced placement and reassigned the same values in lower recall overall for the negative conditions Item 4 as had the corresponding feature in and an interaction between the levels effect Item 3. Items 5 and 6 were considered to be and the type of response. The pattern pro- the context and target in the category-negative duced by the model was due to the similarity condition; Items 7 and 8 were considered to values among the items, and the fact that be context and target in the rhyme-negative given the encoding and retrieval mechanisms condition. These four items were left in their in CHARM, the similarities of the items change original unrelated form. the results of retrieval. The composite trace was constructed as The resonance values produced by the follows: simulation are shown in Table 2. It can be / • seen that the target item had the highest (Item 1 *Item 2) + (Item 3* Item 4) resonance value in the high-similarity (cate- + (Item 5* Item 6) + (Item 7* Item 8), gory-positive) condition, and the lowest res- onance value in the unrelated (category- and where * denotes convolution, as before. Only rhyme-negative) conditions. The criterion can the central n (63) features of the convolutions have a radical effect on responding, however. were used in this simulation and in all sim- For instance, the high criterion reduced the ulations in this article. rate of recall on the unrelated conditions To retrieve, each of Items 1, 3, 5, and 7 from about 85% to about 14%. The rate of were correlated with the 63-tuple representing intrusions was also reduced from 11 total in the composite trace, to result in four retrieved the low-criterion simulation, to 1 in the high- items. The retrieved items were then matched criterion simulation. A very low (sub-zero) to every item in the lexicon, except the item criterion would correspond to a forced recall that had been used as the retrieval cue. To situation, in which a subject is not allowed match, the dot product was calculated be- the option of not responding. A very high tween the retrieved items and each lexical criterion could, of course, eliminate all re- item, to result in a resonance value for each sponding even if the appropriate information lexical item (for each of the four cues). The were retrieved from memory. lexical item that had the highest resonance value above a minimum criterion was consid- Discussion ered to be the item that had been recalled. Two criteria were used, 0.5 and 1.0. (The The reason the model produced the results value of 1.0 was chosen because it produces given in Table 2 is straightforward. When the numerical values that are very close to those context item is used as a retrieval cue, in the found in the Moscovitch & Craik, 1976, CHARM model, recall of the target is an experiment. The value of 0.5 was chosen to increasing function of the similarity of the illustrate the effect of lowering the criterion.) context associated with the target, and the The entire simulation was replicated 100 target (so long as the rest of the list items are times with a different randomization for the unrelated to the cue). In the case where lexical representations used for each replica- similarity is high, the retrieval cue recreates tion. What (if anything) was recalled on each not only the target item, but also the context replication for each cue was tabulated, and itself. The extent to which the context is 20 JANET METCALFE EICH Table 2 Resonance and Recall Scores From Simulation 1 Recall criterion Resonance (%) Condition Target Intrusions 0.5 1.0 High similarity (category positive) 99 80 M 1.3 0.0 SD 0.3 0.3 Medium similarity (rhyme positive) 91 33 M 0.9 0.0 SD 0.3 0.3 Unrelated (category and rhyme negative) 85 14 M 0.7 0.0 SD 0.2 0.2 recreated depends on the extent of similarity CAT. A particular target was presented only between the context and the target. The re- once in the list, but all of the target items construction of the context helps recall of the had both a rhyme and a high associate that target by increasing the strength of the features had previously been equated for the ease of that are common to both the context and the generation of the target from semantic mem- target. ory. The a priori probability of producing As shown in the next simulation, it is not the target items in the absence of any exper- only the similarity between the encoded con- imental input into episodic memory was .16 text and the target that is important in for the rhyme cues and .15 for the semantic CHARM, but also, and fundamentally, the cues. These probabilities indicate whether the nature of the cues that are correlated with target was one of the words (usually two to the trace at time of retrieval. four) that were produced within a 10-s inter- val. Thus, the a priori likelihood of saying Encoding Specificity and CHARM cat as a response to "rhymes with hat" was about the same as the a priori likelihood of Tulving has proposed that factors such as saying cat to "is associated with dog." At depth of encoding, distinctiveness, and elab- time of test, half of the semantically encoded oration need not be considered separately targets were cued by the items with which from or in addition to the compatibility they had been encoded (compatible cueing relation between the trace and the cue, that conditions), whereas half were cued with the is, that the idea of encoding specificity (Tul- appropriate extralist rhyme cue. For the ving, 1979; Tulving & Osier, 1968; Tulving rhyme encoded items, half were cued with & Thomson, 1973) applies to levels data. An the studied (compatible) rhyme cue and half experiment by Fisher and Craik (1977, Ex- with the extralist associate cue. The results periment 2) provides an interesting case in are shown in the left panel of Table 3. There point because it showed an encoding specific- was a levels-of-processing effect—semantic ity interaction and also a main effect of levels better than rhyme—but it was obtained only of processing. when the list cues were given at time of test. Fisher and Craik (1977) presented subjects The finding of a main effect of semantic with a list of 62 pairs of words. Four pairs over rhyme, in the compatible cuing condi- made up a primacy buffer and four were tions, is consistent with the idea that the included in a recency buffer that were ex- semantic encoding led to a deeper (richer, cluded from further consideration. Of the 54 broader, more elaborate, or more distinctive) remaining pairs, half were constructed such trace. However, none of the constructs that that the cue was a high associate of the target, have been used to account for the better as in dog-CAT, and half were pairs in which performance with semantic processing are the cue rhymed with the target, as in hat- sufficient to explain why this superior seman- LEVELS AND CHARM 21 tic trace was not also shown to produce avoid a circularity problem that is implicit superior recall when cued with the incom- in the compatibility explanation. In particular, patible cues. If the trace was richer in the the compatibility explanation allows no semantic condition, this richness had no im- method independent of the to-be-explained pact when the extralist cues were used. recall for assessing what will or will not be a Tulving has argued that Fisher and Craik's compatible relation. (Note that T. Nelson, (1977) data are most parsimoniously ex- 1977, leveled a similar criticism against the plained in terms of only the compatibility depth of processing idea.) CHARM allows that relation between the cue and the trace. A the similarity among items may be ascertained retrieval cue is proposed to be effective only independently of recall. to the extent that its informational content In Simulation 1, it was assumed that the matches, overlaps, or is compatible with the cue and the target were more similar to one information content of the trace. In CHARM, another in the semantic (positive) condition the cue and the trace in no sense match or than in the phonemic (positive) condition. It overlap because the meaning of the dimen- was this difference in similarity in combina- sions in the cue and trace are quite different tion with the associative, storage, and retrieval (see the earlier section of the article, on mechanisms of CHARM that explained the Convolution). Theiermcompatibility, however, main effect of depth of processing in Mos- could be liberalized so as not to implicate a covitch and Craik's (1976) experiment. Fisher matching mechanism for retrieval. This would and Craik's (1977) data are explained in the be consistent with the mechanisms in same way in the simulation that follows. CHARM.2 The model, however, is able to Having proposed that the representation of the items—and their similarity to one an- other—is an important factor influencing re- TableS call, it is possible to obtain an independent Proportion Recalled as a Function of Encoding measure of the similarity (at least ordinally) Context and Type of Cue without recourse to the recall data. Thus, the Encoding context model circumvents the circularity implicit in the compatibility explanation. In the experi- Retrieval cue Rhyme Associate ment that follows, the hypothesis—that the semantic pairs in Fisher and Craik's experi- Fisher and Craik (1977) Experiment 2" ment shared more properties in common Rhyme than did the phonemic pairs—is tested. Target recall .26 .17 A priori probability .16 Associate .44 Experiment 1 Target recall .17 A priori probability .15 Subjects were presented with the pairs of items from Fisher and Craik's (1977) exper- Simulation 2 iment and were asked to write down all of Rhyme the properties that the two items had in Target recall .32 .01 common. Note that subjects were not asked Resonance to free associate to the items, but were spe- M .91 .34 cifically asked for ways in which the items SD .27 .26 Associate were the same. The hypothesis of the exper- Target recall .04 .43 iment was that subjects would be able to Resonance produce more properties in common for the M .44 .99 semantic than for the phonemic pairs. SD .32 .33 Method. Eighteen subjects were told that * From "Interaction Between Encoding and Retrieval Op- erations in Cued Recall" by R. P. Fisher and F. I. M. Craik, 1977, Journal of Experimental Psychology: Human 2 Such liberalization of the compatibility idea may be Learning and Memory, 3, p. 707. Copyright 1977 by implicit in Tulving's (1983a) synergistic ecphory proposal. American Psychological Association. Adapted by permis- He (Tulving, 1983b) has recently stated that this view is sion. compatible with the correlation retrieval operation. 22 JANET METCALFE EICH they would see pairs of items for which they were 63 features representing each item. The were to write down all the common properties items were then normalized so that the dot or characteristics shared by the two items in product of an item with itself gave a value of 1. each pair. Four minutes were allowed for Four blocks of items were constructed each pair. Each subject received 12 pairs—6 starting from the lexical representations so of which were phonemic and 6 of which were that the first item in each block was the target semantic. The actual pairs were randomly item, the second was semantically related to assigned with the constraint that no subject the target item, and the third was phonemi- saw the same item in both the phonemic and cally related to the target. Items 1, 4, 7, and semantic condition. The 18 subjects provided 10 were designated as target items, and were two complete replications of all of the pair left in their original form (i.e., they were combinations that were used in Fisher and unrelated to each other). Items 2, 5, 8, and Craik's Experiment 2. An additional 9 sub- 11 were designated as semantically related jects provided a third replication. The subjects items that were modified so that they were in the third replication were given 3, rather somewhat similar to their respective target than 4 min to write down the common items. In order to represent the high similarity, properties. Subjects were asked to write each 36 features were randomly selected without property on a separate line. To score the replacement and reassigned the numerical data, the lines were simply counted; no at- values of the corresponding target item on tempt was made to ferret out redundancies those features. Thus for instance, Item 2 was or good or bad properties. Subjects were given the same feature values as Item 1 on introductory psychology students at the Uni- 36 features. Items 3, 6, 9, and 12 were versity of California, Los Angeles. They were designated as phonemically related to the run in groups. target items, and were assumed to have low Result. The mean number of common similarity to the targets. Twenty-seven features properties given for the six semantic pairs was were randomly selected without replacement 42.0 and for the six phonemic pairs, 29.4. This and reassigned the same value as had the difference was significant, t(26) = 4.83. corresponding target items on those features. The net result of this reassignment of Simulation 2 feature values was that the semantic items were more similar to their targets than were In this simulation, the representational as- the phonemic items. A less obvious result sumption, consistent with the results of Ex- was that there was a small amount of simi- periment 1, was made that the semantic pairs larity induced between the semantic and pho- in Fisher and Craik's (1977) experiment con- nemic items within a block, by virtue of the tained more features in common than did fact that they were both similar to the,target. the phonemic pairs. It was further assumed, By chance, occasionally some of the features as in other applications of the model, that selected to be overlapping for the phonemic when subjects encoded a pair of items, the items were also among the overlapping fea- items were associated by convolution, stored tures in the semantic items. in a composite memory trace, and retrieved The composite trace was contructed as by correlation of the cue with the trace. The retrieved item was then designated as a par- (Item 1 * Item 2) + (Item 4* Item 5) ticular response by choosing the lexical item + (Item 7* Item 9) + (Item 10*Item 12). that had the highest resonance score with the retrieved item, above a lower criterion. As in The items in the first two pairs were seman- the first simulation, the cue item was not tically related, whereas those in the last two considered as a response possibility. pairs were phonemically related. (Convolution Method. A lexicon of 18 items was con- is symmetric and so it does not matter structed. All 18 items were first assigned whether Item 1 is convolved with Item 2 or numerical values on each feature by randomly vice versa.) The convolutions were truncated selecting each feature value for each lexical to the central 63 features, and so the com- item from a unit normal distribution. There posite trace was a 63-tuple. LEVELS AND CHARM 23 To mimic retrieval corresponding to Fisher CAT depends mainly on the similarity of hat and Craik's (1977) Experiment 2, an intralist and dog. But dog is also retrieved to the semantic cue (Item 2—target is Item 1), an extent that hat and CAT are similar, in this extralist phonemic cue (Item 6—target is situation. If hat-CAT is encoded and dog is Item 4), an intralist phonemic cue (Item 9— given as a cue, the retrieval of CAT once target is Item 7), and an extralist semantic again depends mainly on the similarity of cue (Item 11—target is Item 10) were sepa- hat and dog. However, hat is also retrieved rately correlated with the composite trace to to the extent that dog and CAT are similar. produce four retrieved items. Each of the There is a trade-off at work with the second four retrieved items was then matched to (nontarget) item that is retrieved in the two each item in the lexicon, except the cue itself, extralist cueing cases. In the first case, the to produce resonance scores for the lexical second item, dog, is not retrieved as strongly items. The lexical item with the highest res- as is the second item, hat, in the second case. onance value (above a lower criterion of 1.0) However, though not retrieved as strongly, was selected as the response. The entire sim- dog has more features in common with the ulation was replicated 100 times. target CAT than does hat, and so its retrieval Results. The top panel of Table 3 shows has a greater tendency to help identification the results of Fisher and Craik's (1977) Ex- of the retrieved item as being CAT. It is periment 2; the bottom panel gives the results conceivable that these predictions might be from Simulation 2. It can be seen that, in testable. The high level of recall with the the simulation as well as in the data, there noncompatible cues that was found in the was a main effect of levels of processing. This data (.17) but not the model may be due, to main effect, however, was found only in the some extent at least, to the a priori (nonepi- conditions in which the retrieval cue is the sodic) probability of the responses. This point context with which the target item was asso- is discussed in more detail in a section that ciated. In those conditions in which an ex- follows. tralist retrieval cue was used, there was no levels effect in either the simulations or the Elaboration and CHARM experiment. In the simulation, recall was low A number of investigators (J. R. Anderson in the noncompatible cueing condition. An & Reder, 1979; Battig, 1979; Craik & Tulving, encoding specificity interaction with levels of 1975; Moscovitch & Craik, 1976; Ross, 1981) processing was shown by both the simulation have proposed that levels of processing effects and the data. The ordering of conditions may be explained by means of the construct shown in the data: encode semantic-retrieve of elaboration. In the Fisher and Craik (1977) semantic > encode phonemic-retrieve pho- experiment, for instance, it would be argued nemic > encode semantic-retrieve phone- that when subjects are asked to process hat- mic = encode phonemic-retrieve semantic, CAT, they form fewer other associations to was reproduced by the simulation. CAT than when they are asked to process dog-CAfT. This interpretation of the term Discussion elaboration has been most explicitly formal- ized by J. R. Anderson and Reder (1979). A The reason the model produced the levels larger number of potential routes to the target effect in the compatible cueing conditions is item is assumed to be a concomitant of the same as has already been discussed for semantic processing and is assumed to facil- Simulation 1, and in the description of the itate recall. model. The lack of difference and low level An important question to ask, with respect of recall in the noncompatible conditions to this explanation of levels effects, is whether may be understood from Equation 5. Al- having more elaborations necessarily helps though the level of recall is low, qualitative recall. The answer to this question is no. differences in what is predicted to be recalled There are several studies that show that elab- in the two noncompatible conditions may be oration may or may not facilitate recall. interesting. If dog and CAT are associated Fisher and Craik (1980) have shown a facili- and hat is given as a cue, the retrieval of tative effect, but only under highly circum- 24 JANET METCALFE EICH scribed conditions. Elaboration was denned, orated text (Reder, 1982). As Reder detailed, in their study, as the complexity of the sen- unelaborated summaries produced better tence in which a to-be-remembered word was performance regardless of (a) whether direct embedded. When, at test, the target word was or indirect questions were asked about the presented alone, there was no effect of elab- stories; (b) a delay in testing (up to one year); oration. Their second experiment further (c) the overall study time being greater for qualified the elaboration effect, showing that the elaborated conditions; (d) the nature of even when highly elaborative sentence frames the test being transfer to the learning of new were presented at time of test, a facilitative material, or yes/nor questions, or cued recall effect was obtained only when the elaborations (Reder & Anderson, 1982), or free recall were of a particular kind. The kind of elab- (Allwood, Wikstrom, & Reder, 1982); (e) oration that did facilitate recognition was whether subjects studied in the lab or in called highly redintegrative, where the deter- more naturalistic settings; or even (f) whether mining characteristic was that the sentences subjects were given credit, in the elaborated contained a number of items that were high condition, for recall of noncentral points that associates of the target items. An example of were not even presented in the unelaborated an elaborative sentence that facilitated rec- condition. ognition was, "He felt clean after washing in The fundamental assumption that underlies a hot BATH." Given the results of Experiment the proposal that levels-of-processing effects 1, it would appear that the high associates are really due to more elaboration in the shared many features with the target items. semantic than in the phonemic conditions, is It appears possible that the nature of the that elaboration helps performance. Some- items as highly similar to the target (and one times elaboration does help performance, but another) was critical. sometimes it hurts! It appears that the con- Bransford, McCarrell, Franks, and Nitsch struct of elaboration is inadequate as an (1977) conducted an experiment in which a explanation of levels effects. Because the me- large number of elaborations resulted in morial consequences of elaboration vary, it poorer recall than did a smaller number of appears rather that some explanation is re- elaborations. In the experiment, the minimal quired for the effects of elaboration. elaboration condition consisted of pairs such An experiment by Bradshaw and Anderson as "whale-skyscraper (large), mosquito-doc- (1982) is of particular interest and is modeled tor (draw blood), and lamb-blanket (wool)," shortly with CHARM. Bradshaw and Anderson and the multiple-elaboration condition con- (1982) found both facilitative and inhibitory sisted of elaborated pairs such as "whale- effects of elaboration in the same experiment. deer (eat, move, sleep), mosquito-racoon In the experiment, the extent as well as the (head, legs, jaws), and lamb-snake (breathe, kind of elaboration was manipulated. Three exist, eventually die)." The similarity relations conditions are considered. In the unelabo- of the items to each and every item in the rated-control condition, subjects were pre- list are reasonably complex in this experiment. sented with sentences such as "NEWTON be- The main point for the purpose of the present came emotionally unstable and insecure as a argument—that elaboration does not neces- child." At time of test, the predicate ("became sarily help—is that in the more elaborated emotionally unstable and insecure as a child") condition, recall was poorer than in the less was presented and recall of the subject (NEW- elaborated condition. TON) was required. In the elaborated condi- Reder and Anderson (1980a), in investi- tions, participants were provided with the gating whether elaboration helped recall, rea- same core sentence, and the retrieval cues soned that if elaboration helped, recall for were the same sentence fragments as in the the main points of a text would be superior control condition. In addition, though, they when the fully elaborated text was presented, were provided with other statements concern- as opposed to when an unelaborated summary ing the target. In what is called the similar- of the same major points was given. However, elaborated condition, participants were shown it was found that summaries of texts produced sentences such as "NEWTON became irration- performance better than did the original elab- ally paranoid when challenged by colleagues," LEVELS AND CHARM 25 and "NEWTON had a nervous breakdown ture that specifies the difference between the when his mother died." In the unrelated- similar- and unrelated-elaborated conditions elaborated condition, sentences such as (Reder & Anderson, 1980b) are unelaborated- "NEWTON was appointed warden of London control = unrelated-elaborated > similar- mint," and "NEWTON went to Trinity College elaborated. In contrast to these predictions in Cambridge," were given in addition to the the results of the experiment were similar- core sentence. elaborated > unelaborated-control > unre- Before turning to the data, let us consider lated-elaborated. some predictions. If it is the case that the number of elaborations determines recall and Simulation 3 that the more elaborations there are, the The essential relationships set up by the better is recall, then there should be no experimenters in Bradshaw and Anderson's difference between the similar-elaborated and (1982) experiment may be denoted by con- the unrelated-elaborated conditions. This sidering each sentence to be a subject-pred- seems to be the result expected from J. R. icate pair. Using interference-theory notation, Anderson and Reder's (1979) and J. R. An- in the similar-elaborated condition, the list derson's (1983) model. The pure-elabora- was of the form A-B, A-B'; in the unrelated- tion hypothesis indicates that the results elaborated condition it was of the form A-B, should be similar-elaborated = unrelated- A-C; in the unelaborated-control condition elaborated > unelaborated-control. it was simply A-B. At time of test, B was A prepositional network representation of provided as a retrieval cue in all three con- the difference between similar and unrelated ditions, and recall of A was required. elaborations has been given, in a different In the simulations that follow, I show the context, by Reder and Anderson (1980b). It results produced by CHARM when the simi- was proposed that when the predicates con- larity of the elaborator to the cue varies and nected to the same subject are related (i.e., also when the number of elaborators varies. similar), an integrative subnode is formed, Elaborators are either similar or unrelated to intervening between the subject and the pred- the cue. Either one or two elaborators are icates that are similar to one another. This used. Because Bradshaw and Anderson (1982) particular location of the subnode was re- included several delay conditions, the simu- quired so the model could account for other lations give an imitation of delay as well. data that are not of concern here. When the Method. A lexicon of 18 items was con- predicates are unrelated, no intervening sub- structed by randomly selecting feature values node is formed. According to the preposi- for each item from a unit normal distribution. tional network view of retrieval, search begins There were 63 features in each item. The from the terminal nodes of the fragment vectors were then normalized so that the dot presented as a cue, and spreads along the product of an item with itself was 1.0. experimentally marked pathways. If an inte- The items that would serve as the similar grative subnode exists between the cue pred- elaborators were constructed. Items 3 and 4 icate and the subject (target) in the similar- were reassigned feature values on 36 randomly elaborated condition, search should spread selected features so that they overlapped with from the cue to the subnode and from thence the values of Item 2. Items 9 and 10 were to both the similar elaborations and the target. reassigned values of 36 features so that they In short, a fan effect should occur at the were the same as Item 8 on those features. locus of the integrative subnode-decreasing The traces for the unelaborated-control, the the likelihood of recall of the target. Because unrelated-elaborated, and the similar-elabo- there is no postulated subnode between the rated conditions were constructed as follows: cue predicate and the target in the unrelated- unelaborated control (Item 1 *Item 2) + (Item elaborated condition, there should be no in- 7* Item 8); unrelated-elaborated (1) (Item tervening fan effect to impair recall, and 1 *Item 2) + (Item 1 *Item 5) + (Item 7* Item performance should be the same as would be 8) + (Item 7* Item 11); similar-elaborated (1) the case had there been no elaboration. The (Item 1 *Item 2) + (Item 1 *Item 3) + (Item predictions of a prepositional network struc- 7* Item 8) + (Item 7* Item 9). 26 JANET METCALFE EICH The extent of elaboration was also varied B'2 (Item 4) all served to reconstruct the by using two elaborations for each target Target A (Item 1). In the unrelated-elaborated rather than just one. The unrelated-control condition, the unrelated cues (Item 5 and 6) condition was the same as above: (Item had no features in common, beyond indepen- 1 *Item 2) + (Item 7* Item 8). The unrelated- dence, with the retrieval cue, and so no signal elaborated (2) trace was (Item l*Item 2) + components were systematically extracted (Item 1 *Item 5) + (Item 1 * Item 6) + (Item from those associations. However, the unre- 7 * Item 8) + (Item 7 * Item 11) + (Item lated associations did produce noise compo- 7* Item 12). The similar-elaborated (2) trace nents not present in the unelaborated-control was (Item l*Item 2) + (Item Ultem 3) + condition. The additional noise in the unre- (Item 1 *Item 4) + (Item 7* Item 8) + (Item lated-elaborated condition accounts for the 7* Item 9) + (Item 7* Item 10). poor recall in that condition. Delay was mimicked by adding unrelated noise—(Item 17*Item 18)X4—to each of Parameter Values the traces detailed above. Thus, in the une- I discuss here the reasons for the choices laborated-control-delay condition, for exam- of parameter values used in the above simu- ple, the trace was (Item l*Item 2) + (Item 7*Item 8) + 4(Item 17*Item 18). The extra- Table 4 neous noise was multiplied by 4 so that the Recall Depending on Extent and Kind effects would be fairly obvious. of Elaboration To retrieve, in all conditions, Item 2 was correlated with the trace. The retrieved item Condition Experiment was matched, as in the previous simulations, time Similar Unrelated Control to all items in the lexicon except for the cue Item 2, and the item that showed the highest Bradshaw and Anderson (1982)1 resonance score above a criterion of 0.5 was 1 (immediate) 96% 80% 93% designated the recalled item. The simulations 1 (delay) 92% 74% 87% were replicated 100 times, and mean reso- 2 (delay) 75% 45% 61% nances, standard deviations, and recall fre- 3 (delay) 61% 32% 38% quences were calculated. Simulation 3 (one elaborator) Results. Bradshaw and Anderson's (1982) results are shown in the top panel of Table Immediate Recall 99% 86% 94% 4. The center panel shows the results of Resonance Simulation 3 when one elaborating pair was M 1.20 0.75 0.77 used, and the bottom panel shows the results SD 0.46 0.22 0.18 when two elaborators were used. The pattern Delay produced by the simulation mimics the pat- Recall 69% 48% 53% Resonance tern produced by people. In the simulation, M 1.20 0.76 0.77 as in the data, the unrelated elaborators SD 0.60 0.44 0.43 impaired recall, and the similar elaborators improved recall, relative to the unelaborated Simulation 3 (two elaborators) control condition. Comparison of the center Immediate and bottom panels of the table shows that as Recall 100% 70% 91% the number of elaborators increase, the dif- Resonance M 1.60 0.71 0.74 ferences between the elaborated conditions SD 0.59 0.33 0.17 and the control condition also increase, ac- Delay cording to the model. Recall 88% 42% 49% Resonance M 1.60 0.70 0.73 Discussion SD 0.71 0.53 0.43 The similar-elaborated (2) condition in the * From "Elaborative Encoding as an Explanation of Levels of Processing" by G. H. Bradshaw and J. R. Anderson, simulation produced the best recall because 1982, Journal of Verbal Learning and Verbal Behavior, the components extracted by the Cue B as 21, p. 171. Copyright 1982 by Academic Press, Inc. well as those extracted by B'! (Item 3) and Adapted by permission. LEVELS AND CHARM 27 lations and some of the problems involved itself), in the similarities of items to one in constructing a realistic simulation. I have another, in the magnitude of particular fea- tried to either keep the values of irrelevant tures, perhaps in intercorrelations among fea- parameters constant, to show that the effects tures, and so on. of interest do not interact with the values of The model is sensitive to the number of certain parameters or to show how effects features in the vectors. More features produce vary with variations in parameters. The sim- better recall. If one keeps the dot product of ulations are minidemonstrations of the effects an item with itself constant, then increasing of changing one parameter in particular— the number of features results in each feature the similarity among items. The intent of the on average having a smaller value. The overall article is to demonstrate the changes in this variance is decreased, and increased accuracy parameter produce interesting levels effects results. The same net result occurs, however, that are often empirically observed. The or- if one lets the dot product increase along dinal differences in the similarity parameter with the number of features but keeps the should be open to experimental verification average absolute value of the features constant. (as in Experiment 1). The magnitude of the In this case, the mean dot product of the differences in similarity among conditions is retrieved item with its lexical representation problematic, though, both because of the increases producing increased accuracy of measurement problem involved in relating recall. The number of features per item that subjects' ratings of similarity to the features I used in the simulations was 63 and was in the model and also because the same held constant. This number was arbitrary— results can be produced in the model with chosen only because I had arbitrarily chosen different numbers or proportions of common this number in some earlier simulations (Eich, features (holding the ordinal relations the 1982). The number may be large enough to same) if other parameters are varied, such as suggest that the model is intended to deal the criterion, or the number of associations with large numbers of features rather than in the memory trace, or the number of the small numbers used in some feature features in the items. Although a number of models. I actually also ran these simulations parameters have implications for recall in the with 21 features per item as well. Qualitatively model, the present article primarily depicts a the results are the same as those reported continuation of an exploration of the reper- here although the variability was greater. A cussions of manipulating one particular pa- guess at the number of features that a realistic rameter—similarity. (rather than a demonstration) simulation There is a real problem in attempting to should use would be a number of the order ascertain appropriate parameter values in the of magnitude of the number of neurons in, simulation model because the model is sen- say, the hippocampus. Figure 1 in Metcalfe sitive to factors and variables that are im- and Murdock's (1981) article illustrates the practical to simulate. For example, the model linear increase in the level of recall that is sensitive to the size of the lexicon. As obtains as the number of features is increased. lexicon size increases, recall falls off. The The work of Tversky (1977) suggests that it maximum size lexicon that I have simulated might be reasonable, indeed necessary, to was 350 sixty-three feature items. All of the represent different items as being coded on items were (unrealistically) random vectors. different numbers of features (the uncoded The pattern of results that was obtained with features could be assigned a value of zero). these simulations mirrored the pattern found The implications for memory of varying the with a small lexicon of 50 items, except that number of features, across items, could be the level of recall was roughly 20% lower (see but have not yet been investigated within the Metcalfe & Murdock, 1981, Figure 2). To CHARM model. realistically imitate human recall, one would The dot product of an item with itself was like to build in a lexicon of several thousand always set to 1.0 in the simulations. Changes items (words). Of course, one would not want in this value might reflect differences in effort those items to be represented as random and arousal. Although it is well known that vectors. Rather, they should vary in a priori arousal manipulations may affect memory strength (the dot product of an item with even within the context of levels experiments 28 JANET METCALFE EICH (Krinsky & Nelson, 1981), I have not consid- reasons, the numerical values of the param- ered them in the present simulations. eters that were varied—the similarity of the The number of convolutions, or associa- items to one another and the value of the tions, that are added into the composite trace response criterion—may be different from influences the level of recall. If all of the those that would be used in a more compre- items are unrelated and represented as ran- hensive simulation. dom vectors with a mean of zero, then recall decreases monotonically as a function of the Simulation 1 number of pairs. In particular, if nothing is Having done a considerable amount of done to control the variability of the com- hand waving about why the parameter values posite trace, then the expected mean dot should not be taken too seriously and that product of the retrieved item with its correct the pattern of effects shown by the model, lexical entry stays the same, but the variance but not the absolute levels of recall, are increases linearly with additional associations. important, I will now precede to admit that If the variance is controlled (by say renor- I put some effort, in Simulation 1, into malizing the trace with each entry, or by ascertaining values of similarity and a value assigning appropriate weights to each new of the response criterion that would produce entry and to the trace, or by allowing decay, numbers that correspond to those found in attenuation, or inhibition to occur in the Moscovitch and Craik's (1976) experiment. model), then the decrease in the level of recall The major reason for fitting the data from still occurs as more and more entries are this experiment quite closely is that a casual added into the trace. However, in this case inspection of Equation 5 might suggest that the poorer performance would be attributed differences in the level of recall that exceed to the mean dot product rather than to the a 2:1 ratio could have been impossible to variance. produce with the CHARM model. Moscovitch There should probably be a parameter that and Craik's results exceed the 2:1 ratio. How- gives the probability that a given association ever, as the first simulation shows, the model was formed and stored at all. There are many can be fit to these data fairly readily. To do precedents for assigning a probability value so, I used a high-response criterion. This to the entire process that is modeled by threshold was treated as a free parameter, convolution and addition in the present and the results of lowering it are shown in model. If a subject's attention wanders, for Table 2. The high threshold has the additional example, the to-be-remembered event may benefit over the low threshold, of almost not be processed and stored in episodic mem- completely eliminating intrusion errors. Peo- ory. There may also be neurological insults ple do not typically produce many intrusions, that affect the entire associative-storage so this is a pleasing spin-off of the high mechanism. I do not rely at all on variations threshold. in the probability of association formation and storage or on the probability of enacting Simulation 2 the retrieval process in any of the simulations that I have conducted on the CHARM model. In Simulation 2 I left the similarity param- This potential parameter is set to 1.0 in the eter of the rhyme condition at the same value model and so failure to store or to go through it had had in the first simulation. There was the retrieval process is not responsible for about a 5% difference in the level of recall in any of the effects that are demonstrated by this same condition from Moscovitch and the model in this article. Craik's (1976) to Fisher and Craik's (1977) In the demonstration simulations that have experiment. I do not know why this difference just been presented, I did not set up structured occurs in the two experiments, or whether it lexicons of 10,000 items or use vectors with is a meaningful difference. The materials 106 dimensions or even use the same number were no doubt somewhat different, and it of items as were used in the experiments. I may have been the case that Fisher and used only enough associations to exemplify Craik's rhyming pairs were slightly more the designs of the experiments. For these similar to one another than were Moskovitch LEVELS AND CHARM 29 and Craik's rhymes. It would have been very associative question and not just the sole easy to modify the value of the similarity response. It seems plausible to speculate that parameter in the second experiment to pro- the . 17 levels of responding might have re- duce a more precise result in the rhyme- sulted from some combination of guessing rhyme condition, but I saw no clear justifi- and semantic priming. The semantic priming cation for doing so. On the other hand, there could have occurred because the target word was no logical reason for making the free- had been recently presented (but in a different association condition in the Fisher and Craik context). (1977) experiment equivalent to the category condition in Moscovitch and Craik's (1976) Simulation 3 experiment. In this simulation I attempted to show that What the model does not do is almost as the model would produce the right pattern interesting as what it does, in the situation of results and that increasing the extent of set up by Fisher and Craik (1977). The levels the elaboration manipulation, by increasing of recall in the noncompatible-cuing condi- the number of pairs, would exaggerate the tions were .17 and .17. As can be seen from effects, both positive and negative. Because Table 3, the simulation produces a level the conditions in this experiment were quite between .01 and .04—considerably lower than different from those in the previous two that shown by subjects. A retrieval cue that experiments, it did not seem necessary to is not similar to the cue associated with the maintain constant parameter values across target simply does not retrieve the target item simulations. in the model. Manipulating parameters does It may be interesting that the same pattern not substantively change this result in the of facilitation and inhibition would result in model. Unless the similarity between the the situation set up by Bradshaw and Ander- encoded cue and the retrieval cue is quite son (1982) even if the events were represented high, poor target recall results. Because the somewhat differently. If each proposition were model does not produce a high level of target treated as a vector and the target were treated recall to the extralist cue, it is necessary to as part of that vector, then one may perform claim that the high (.17) level of correct an autoconvolution (A* A) at time of storage. responding in the noncompatible-cuing con- At time of retrieval a fragment of A (the ditions in Fisher and Craik's (1977) experi- predicate) suffices to reconstruct all of A ment is due to some mechanism other than including the target, by correlation. The episodic retrieval from the composite memory goodness of reconstruction of the missing trace. Alternatively, the mechanisms in the target fragment will be a function of the model might be wrong. Fortunately for the similarity of the elaborating propositions. The model, it is defensible to claim that the possibility of autoconvolution as an encoding responding in the noncompatible-cueing con- operation is pursued in more detail in the ditions is not purely episodic, and that the section that follows on recognition memory. model, which is currently explicit only about episodic retrieval, should not account for the Recognition and CHARM high level of responding. In particular, Fisher Many experiments within the levels-of- and Craik (1977) determined the a priori processing framework have used recognition level of response production under conditions as the memory test, rather than recall. Rec- where there was no episodic input at all. This ognition experiments, like recall experiments, level was . 15 in the associative-cue case, and show sensitivity to levels manipulations (e.g., .16 for the rhyme cues. These levels are very Craik & Tulving, 1975; Jacoby & Dallas, close to the .17 levels shown in the experi- 1981). Several researchers have shown that ment. However, there is probably something the preexperimental associative value of other than pure guessing occurring to produce paired items (which, as Experiment 1 shows, the . 17 levels found in the experiment because is confounded with similarity) influences rec- the .15 and .16 levels refer to the probability ognition (Kinsbourne & George, 1974; D. that the target is one of the two to four words Nelson, Brooks, & Wheeler, 1975; Rosenberg, produced as a response to the rhyme or 1968; Underwood, 1976), although these re- 30 JANET METCALFE EICH suits are not always straightforward (e.g., occurs; if the probe does not retrieve itself, Fisher, 1979). In this section, I give a thumb- then recognition does not occur. If an inter- nail sketch of a composite holographic au- item association as well as an autoassociation toassociative recognition model, and show was formed, the trace also supports recall or that the recognition model produces changes a recall check in a recognition task. I deal in recognition depending upon context. My only with the more straightforward recogni- intention is only to show that some standard tion process here. The mechanics of retrieval levels of processing effects can be handled in in recognition are identical to those in recall. recognition within the CHARM model. A more The decision process, however, is different. In thorough and encompassing exposition of recognition, once an item has been retrieved, recognition may be presented in a future the question is whether the retrieved item is article. the same as or different from the probe. If The CHARM recognition model outlined the probe item has been present in the list, here had its inception in a conversation with then its autoassociation should have been R. M. Shiffrin (personal communication, added into the composite trace. If the au- February 22, 1983), and bears a resemblance toassociation was added into the trace, then in some respects to the recognition model of correlating the probe with the trace produces Gillund and Shiffrin (1984), as well as to the an approximation to the probe item (plus holographic recognition model of Cavanagh noise, and possibly plus other items that may (1976). Following these theorists, it is assumed be more or less similar to the probe). If the that recognition memory is, to some large autoassociation was not added into the trace, extent, based on the autoassociation of the the correlation of the probe with the trace list items. The model bears a resemblance to should (if the similarity of the probe to the those of Mandler (1980) and Murdock (1982) list items is zero) produce only noise. The insofar as two kinds of information are im- decision process, therefore, assesses whether plicated. However, neither Mandler's nor the item retrieved from memory is similar to Murdock's recognition model use the con- or different from the probe item. volution-correlation storage and retrieval op- To decide whether the retrieved item is erations. The model differs radically from like the probe, the central n features of the Mandler's (1980) recognition model insofar retrieved item are compared with the features as Mandler proposes a search as the retrieval of the probe. If the product of a Feature f of process. The idea of searching is incompatible the retrieved item and Feature f of the probe with the retrieval mechanisms in CHARM. is a positive value, the item can be said to Autoassociation is easily written in the match on this feature and a positive accu- CHARM model as autoconvolution (i.e., A* A). mulator is incremented by the value of the Suppose that a subject is presented an A-B product. If the product is negative, the features pair. To handle recognition as well as recall, are contrasting, and a negative evidence ac- it is assumed that not only is A convolved cumulator is incremented by the value of the with B, but also that A is convolved with product. The positive and negative incremen- itself, and that B is convolved with itself. The ters race to positive and negative criteria and trace, T, is (A*B) + (A*A) + (B*B). Very the one that reaches its criterion first deter- likely, there should be parameters attached to mines whether the response is yes or no. the probability of auto and interitem convo- Decision processes, similar to this one, have lution, but they are not discussed here. It been used previously in recognition memory may be noted that the addition of autoasso- models (J. A. Anderson, 1973; Ratcliff, 1978; ciations in the composite trace does not and see also Link, 1975; T. Wickens, 1982). change the applicability of Equation 5, except To obtain reasonable results, the criterion that the autoconvolutions must be included, for the positive accumulator (for a yes re- in T, of course. sponse) must have a greater absolute value It is assumed that when a probe is presented than the criterion for the negative accumulator in recognition, the probe is correlated with (for a no response). This requirement results the trace to produce a retrieved item. If the from the fact that the expected dot product probe retrieves itself, then positive recognition between the item retrieved by an extralist LEVELS AND CHARM 31 probe and the probe is zero. If the yes and Table 5 no criteria were the same, the probability Results of Simulation 4a: Recognition would be 0.5 that a new unrelated item would incorrectly be called old. In a similar Responses (%) manner, if only one accumulator were used, Condition Yes No in which both positive and negative values were added, the model would have serious Old item High-similarity pair 100 0 problems in accounting for no, and especially Low-similarity pair 88 12 fast no responses, because, even for unrelated Unrelated pair 71 29 items that were not in the list, the expected New item (Unrelated) 17 83 dot product is zero, rather than a negative value. condition; Item 5 is the probe for the rhyme- Simulation 4a or category-negative condition; and Item 7 is an unrelated extralist lure that should be The point of this simulation was to show given a no response. Each of these four that a main eifect of level of processing would probes retrieved a pattern that was then obtain in recognition memory with the sim- compared to the probe that had been used ilarity of the representations coded (as indi- to produce the retrieved pattern. cated earlier in the article), and with the There were two accumulators, one for the encoding, storage, retrieval, and decision pro- yes and one for the no responses. The features cesses outlined above. The simulation de- in the probe and the retrieved item were scribed below is somewhat like Simulation 1 compared serially (presumably with each fea- insofar as three conditions are used: highly ture comparison requiring some unit of time). similar (category positive), moderately similar If the product of Feature f in the probe and (rhyme positive), and unrelated (category or Feature f in the retrieved item was positive, rhyme negative). In addition to each context- the yes accumulator was incremented by the target association, each item is also convolved value of the product. If the product was with itself and entered into the composite negative, the no accumulator was incremented trace. by the absolute value of the product. The Method. A lexicon of eighteen 63-feature criterion of the yes accumulator was set at items was constructed by randomly selecting 0.5, whereas the no criterion was set at 0.25. feature values for each item from a unit The response was determined by the accu- normal distribution, and then normalizing mulator that reached its criterion first. The each item so that F • F = 1. A highly similar simulation was replicated 100 times. The pair was constructed by randomly selecting number of yes and no responses for each of 54 features and replacing the values in Item the four probes was tallied, as were the num- 2 with those in Item 1. Two items, moderately ber of features compared for each response. similar to one another, were constructed by Results. The frequency of yes and no randomly selecting 27 features and replacing responses for each probe is shown in Table in Item 4, the values in Item 3. 5. It can be seen that the items that were The trace was (Item l*Item 2) + (Item members of high-similarity pairs (i.e., deeply 3* Item 4) + (Item 5* Item 6) + (Item 1 *Item processed items) were recognized better than 1) + (Item 2* Item 2) + (Item 3* Item 3) + were items that were members of the low- (Item 4* Item 4) + (Item 5* Item 5) + (Item similarity pairs (i.e., shallowly processed 6* Item 6). Thus, as well as interitem asso- items). The lures were usually correctly re- ciations, as in the previous simulations, in jected. This pattern of results corresponds to this simulation each item was also autoasso- experimental results. It might be possible, in ciated. some future development of the recognition To retrieve, Items 1, 3, 5, and 7 were model, to relate the number of features com- correlated with the trace. Item 1 is the probe pared to reaction time (as in J. A. Anderson's, for the semantic or category-positive condi- 1973, model). There were a considerable tion; Item 3 for the shallow rhyme-positive number of no responses with relatively few 32 JANET METCALFE EICH feature comparisons (i.e., fast noes), so this information that supports recall. The most kind of response does not appear to be a important reason for preferring the present particular problem. scheme for recognition, over Murdock's pro- posal, however, is that the present model can A Comparison to Murdoch's account for the levels of processing data in Recognition Model recognition, whereas Murdock's recognition model does not produce these results. In The recognition system that I have pro- particular, Murdock's model, as it currently posed here is different from that proposed by stands, or even if similarity were represented, Murdock (1982, as well as by J. A. Anderson, could not distinguish between an orienting 1973). Murdock proposed that storage for question that is directed to the correct re- later recognition conforms roughly to a sponse and one that is directed to the wrong matched-filter model. No convolution of the response. Consider List A, category animal- to-be-recognized items is conducted. Further- sheep? category flower-tulip? as compared to more, recognition is not a retrieval process List B, category animal-tulip? category using the retrieval operation of correlation, flower-sheep? If everything were appropri- as I propose here, but rather consists of taking ately counterbalanced, Murdock's recognition the dot product between the probe and the model predicts the same level of recognition trace. In the recognition model proposed here for both lists whereas the CHARM model a representation that is an entity like the predicts better recognition in the first than input items is retrieved from memory. No- in the second list. In short, the associations where in Murdock's recognition model is an make no difference (except for the addition item produced that is itself a vector like the of noise) to the recognition judgment in input vectors. Rather, only a spike or a signal Murdock's model. In the experiments where (a scalar) is produced that varies in magnitude this situation has been examined, a difference depending on the similarity of the probe to is found. Because the CHARM recognition the trace. model proposed here has not been extensively I chose to model the autoassociative, re- tested in other situations, I do not yet know trieval-plus-decision scheme for recognition whether this success in accounting for the over the matched-filter scheme used by Mur- data will hold up in general or not. dock for the following reasons. First, the convolution operator can be thought of as a Simulation 4b transformation that converts the items. When items are associated by convolution, the This final simulation was conducted to meaning of the features in the resultant entity demonstrate that the inclusion of the auto- is different from the meaning of the features convolutions in the composite memory trace in the initial items. In Murdock's model, does not qualitatively change the results ob- features in the transform domain—the asso- tained in recall. ciations—are added into the composite Method. The lexicon and the episodic memory trace, just as they are in the CHARM trace in this simulation were constructed in model. However, vectors that have not been exactly the same way as in Simulation 4a. so transformed and that consequently have a The retrieval and identification mechanisms different status and dimensionality are also were the same as in Simulation 1. added into the same trace. These latter vectors Results. The recall, resonance, and stan- support item recognition. The incompatibility dard deviations of the resonance scores are of the associative and item information in shown in Table 6. As can be seen, the recall the same trace makes Murdock's model con- still reflects the levels manipulation, just as it ceptually complicated. It is not clear what did in Simulation 1, which is similar to the the advantages are, if any, to this scheme. present simulation except that the autoasso- The first reason, then, for preferring the ciations were not included. The resonance present proposal over Murdock's is that the scores are much inflated (as are the standard information that supports recognition is of deviations). These increases are due to the the same kind and dimensionality as is the autoassociations. However, the ordinal pattern LEVELS AND CHARM 33 Table 6 getting from one layer, state, or stage, to Simulated Recall Using the Recognition Memory another. CHARM is thus very much a stage Trace From Simulation 4 theory of memory. It seems to be in the same Type of pair M SD general class as the stage theories Craik and Lockhart (1972) argued against, and to which High similarity they proposed the depth-of-processing theory Recall 100 as a conceptual alternative. Resonance 3.78 0.93 Low similarity In CHARM it is assumed that a number of Recall 92 sensory and perceptual processes occur before Resonance 2.26 0.89 the event in question reaches the form of Unrelated representation (at the item level) necessary Recall 30 for consciousness and for association forma- Resonance 0.87 0.75 Extralist lure* tion. The preconscious operations are not Intrusion 3 assumed to impact on episodic memory. This Resonance 0.05 0.54 model fits nicely with the empirical studies of Jacoby and Dallas (1981), and Jacoby and • This condition gives an approximation to the intrusion rate in the simulation for an extralist cue retrieving a Witherspoon (1982), although not necessarily particular lexical item. with their interpretation. Craik and Lockhart (1972) placed considerable stress on percep- tual and memorial processes. In CHARM, too, is still the same. The proportion recall results the processes used for associating, storing in look rather like the low criterion version of memory, and retrieving (i.e., convolution, Simulation 1, and as in that simulation, there addition, and correlation) are of prime im- were a good number of intrusions. The inclu- portance. However, it is not considered to be sion of the autoconvolutions in the composite the case in CHARM, as seems to be implied trace does not qualitatively alter the biasing by the depth of processing theory, that all effect due to context, that is found in recall. mental operations are memorial operations. Nor does it qualitatively alter the differences CHARM has nothing to say about many cog- in the goodness of recall that are attributable nitive functions such as mental arithmetic, to this biasing. sentence parsing, syllogism construction, or Simulations 4a and 4b provide a demon- indeed, even the memorial impact of similar- stration that the transformational character- ity, category, or rhyme decision processes. istics of the holographic trace, which give rise CHARM deals only with the operations con- to levels-of-processing effects in recall, could sidered necessary for association formation, also produce those effects in recognition. In storage, and retrieval. My argument has been the next section I discuss the relation of the that for a large number of levels effects, that CHARM model to a number of other con- is all one needs to consider. structs that have been proposed to facilitate The episodic memory system in CHARM is our understanding of levels-of-processing ef- envisaged as a rather specialized faculty. This fects. system may interact with other cognitive sys- tems; the changes and distortions in represen- General Discussion tation that are introduced by the memory Comparison of CHARM to system, and that have been the major concern Other Theoretical Ideas of this article, probably impact on functions other than episodic remembering. Neverthe- Depth of processing. The view proposed less, the episodic memory system can be in the CHARM model is very different from thought of as functionally, and perhaps ana- that proposed by Craik and Lockhart (1972). tomically, distinct. Though holographic mod- In CHARM there is a level or layer of repre- els are in the class of distributed models, this sentation at which events are or may be in no way precludes localization of the asso- available to consciousness; there is another ciation or retrieval functions themselves, as layer that is the (composite) episodic memory Figure 1 perhaps suggests. The CHARM model trace; and, there are denned operations for is compatible wilh function-al or operation-^ 34 JANET METCALFE EICH localization where the operations are convo- sions. The representations of the items in lution and correlation. It is incompatible with combination with the associative, storage, the idea that particular discrete episodic and retrieval mechanisms in CHARM poten- memory traces of individual items or events tially lead to what have been called distinc- are localized in the sense of being stored tiveness effects. separately. Though perceptual analysis is, of Compatibility. The explanation of depth course, assumed by the CHARM model, it is of processing given in the present article not considered to be the case that episodic bears a relation to the idea of cue-trace memory is only an automatic by-product of compatibility (Tulving, 1979), congruence perceptual processing. (Craik & Tulving, 1975; Schulman, 1974), Elaboration. The CHARM model does not and similarity (D. Nelson, 1979). The model implicate elaboration as an explanation of also produces the results subsumed under the levels effects as do prepositional network construct of encoding specificity (Tulving & models. Rather, elaboration, if and when it Thomson, 1973). In agreement with these occurs, is (like rehearsal) considered to be ideas, the nature of the materials and the data requiring explanation (Cavanagh, 1976; context in which the to-be-remembered event Metcalfe & Murdock, 1981). The. question is encoded and retrieved are crucial. The with elaboration, as with recall and rehearsal, CHARM model is firmly footed in this inter- is how people are able to retrieve information active approach to human memory and con- that is not in consciousness, regardless of tributes a mechanism whereby the similarity whether that retrieval occurs at time of study intrinsic in the presented events manifests or test. A quote from Nisbett and Wilson memorial consequences. (1977) helps clarify the point. Although the effects of similarity demon- An example of the confusion of intermediate output with strated in this article were facilitative, simi- process was provided by an acquaintance . . . who was larity in the model does not always produce asked to introspect about the process by which he had improved performance. I do not use similar- just retrieved from memory his mother's maiden name. ity, per se, as an explanation. Rather, the "I know just what the process was," he said. "I first thought of my uncle's last name, and since that happens representation of each item is considered to to be my mother's maiden name, I had the solution." be intrinsic to that item. Without some model This only pushes the process question back a step further, saying how representations interact, or fail to of course, and our acquaintance's answer would appear interact, it would not be possible to predict to reflect a confusion of intermediate results with the what kind of effects similarity should have. process by which the final result was obtained, (p. 255) (It is easy to construct models that do not Distinctiveness. A number of researchers produce the effects shown in the data.) have stressed that the distinctiveness of the CHARM provides an interactive model that to-be-remembered events may influence the produces the results found in the data. Pre- memorability of those events (Craik & Jacoby, dictions are derivable from the model and so 1979; Eysenck, 1979; Klein & Saltz, 1976; it is open to falsification. By providing a Lockhart, Craik, & Jacoby, 1976; D. Nelson, working mechanism, CHARM is able to side- 1979). Distinctiveness is not, in and of itself, step the circularity of the compatibility idea, considered to be an explanation in CHARM. while at the same time maintaining and Though distinctiveness sometimes helps per- formalizing the interactiveness that is its cor- formance, the model predicts that it some- nerstone. times hurts performance. For instance, if the cue and target are distinctively different from Transformation of Information one another, the model predicts worse recall than if the cue and target are less distinctive I have argued in this article that many of and more similar to one another. Because the levels effects that have been reported in associations are stored in a composite mem- the literature can be explained by means of ory trace, the model provides a mechanism the operations used for association formation, whereby the similarity or lack of similarity storage and retrieval in the CHARM model, of items, not only within a pair but also and by the a priori representations of the to- between pairs, can have memorial repercus- be-remembered and context items. These two LEVELS AND CHARM 35 threads were tightly woven during the devel- the implications of these properties as a litmus opment of the argument. The fact that the test to assist in taxonomizing different tasks CHARM model would produce differences in and memory systems. Allowing ad hoc mech- the level of recall depending on interpair and, anisms to be added to elegant unidirectional especially, intrapair similarity led me to look associative mechanisms obscures their proper for situations in which these similarities might predictions, and should, I think, be avoided. be a factor in the observed performance. However, I do not deny the possibility that Many levels of processing experiments, as such a modification could produce Equation well as encoding specificity and elaboration 5. Nor do I deny the more interesting possi- experiments, seem to exemplify these differ- bility that it may be possible to develop more ences. Even though the two threads of the elegant or more empirically defensible sym- argument—the processes postulated by the metrical associative schemes than those used model and the representational structure of in the CHARM model, that might produce the information—are both necessary in some Equation 5 and results like those outlined form, it is possible to tease them apart. In here. this final section of the article, I try to do so. Structure. The nature of the to-be-re- Process. It seems reasonable to suppose membered items is of considerable impor- that there exists a class of models, perhaps tance in the CHARM model. If the thesis of involving different mechanisms, that like this article is correct, it is also crucial for CHARM may produce the results of interest many levels-of-processing, encoding specific- in this article, given the representational con- ity, and elaboration experiments. I have pro- straints that I have argued should be placed posed that the content of what is associated, on the to-be-remembered items. Probably the stored, and retrieved (in combination with most important condition that must be met the processes in the model) may result in by a model in this class is that it produce different levels of recall. Equation 5 or something reasonably close to Just as there may be several different pro- it. The association is bidirectional and sym- cesses that are capable of producing Equation metrical. It may produce either or both of 5, there is no doubt also a class of different the associated items. All of the items that are types of representations that support this produced are added in a single representation. equation. In the CHARM model, feature-set At the present time, to my knowledge, representations are used. These representa- there are no other models of human memory tions are probably the simplest and most (except Murdock's TODAM model, for recall easily understood means of representing but not for recognition) that automatically mental events for which Equation 5 is mean- produce Equation 5. However, it would be ingful. However, it is possible to represent possible to modify some extant memory items more complexly—as matrices, for ex- models to do so.3 For example, reduplicating ample—and still maintain the properties of the directional association that is used in interest. It does not seem possible to represent other models (e.g., J. A. Anderson et al., items as undifferentiated points in multidi- 1977; Hintzman, 1983; Medin & Schaffer, mensional space (as has been done in some 1978) by adding a second operation that node-search models) and still show the facil- stores in memory a backward as well as the itative and inhibitory effects of similarity in forward association, might allow those models a manner consistent with the present formu- to approximate Equation 5.1 do not advocate lation. The CHARM model, and Equation 5, this step because I think that it fundamentally depends for its predictive ability on adding undermines a basic concept that is implicit the values of certain features to result in in the directional models, namely, their di- stronger or weaker feature values that may rectionality. It seems most likely that there influence pattern identification and recall. are some mental processes of types or infor- Addition of feature values is fundamental to mation processing for which a unidirectional association is appropriate and even required. 3 One example of this strategy has been advanced and It would be interesting to use the directional brought to the attention of the author following the versus symmetrical associative properties and writing and acceptance of the present article. 36 JANET METCALFE EICH the model. The concept of addition does not Anderson, J. A. (1977). Neural models with cognitive seem to be meaningful, or at least not mean- implications. In D. LaBerge & S. J. Samuels (Eds.), Basic processes in reading (pp. 27-90). Hillsdale, NJ: ingful in a way that makes predictions about Erlbaum. recall under the representational assumption Anderson, J. A., & Hinton, G. E. (1981). Models of of undifferentiated unitary concept nodes that information processing in the brain. In G. E. Hinton exist as points in multidimensional space. & J. A. Anderson (Eds.), Parallel models of associative Thus, the present explanation of levels-of- memory (pp. 9-48). Hillsdale, NJ: Erlbaum. Anderson, J. A., Silverstein, J. W., Ritz, S. A., & Jones, processing, encoding specificity, and elabo- R. S. (1977). Distinctive features, categorical perception, ration data allows for some flexibility in the and probability learning: Some applications of a neural way items are represented but is incompatible model. Psychological Review, 84, 413-451. with node-search conceptualizations of hu- Anderson, J. R. (1972). FRAN: A simulation model of free recall. In G. H. Bower (Ed.), The psychology of man memory. learning and motivation: Advances in research and theory (Vol. 5, pp. 315-378). New York: Academic Summary Press. Anderson, J. R. (1983). A spreading activation theory of In this article I detailed some implications memory. Journal of Verbal Learning and Verbal Be- of a model of human episodic memory that havior, 22, 261-295. combines and modifies events. The result of Anderson, J. R., & Reder, L. M. (1979). An elaborative processing explanation of depth of processing. In retrieval differs from the item that was orig- L. S. Cermak & F. I. M. Craik (Eds.), Levels of inally encoded. The transformations that oc- processing in human memory (pp. 385-404). Hillsdale, cur depend on the similarities among the to- NJ: Erlbaum. be-remembered items, the association and Atkinson, R. C., & Shiffrin, R. M. (1968). Human memory: A proposed system and its control processes. storage of items by convolution and addition, In K. W. Spence & J. T. Spence (Eds.), The psychology and the representations of the cues that are of learning and motivation: Advances in research and used to retrieve by correlation. One important theory (Vol. 2, pp. 89-195). New York: Academic point to be taken from the model is this: Press. Both the intrinsic similarity among items— Barnes, J. M., & Underwood, B. J. (1959). "Fate" of first-list associations in transfer theory. Journal of the representational structure—and the as- Experimental Psychology, 58, 97-105. sociative operations that subjects perform to Battig, W. F. (1979). The flexibility of human memory. interact items—the cognitive process—are In L. S. Cermak & F. I. M. Craik (Eds.), Levels of important and neither of these two ideas is processing in human memory (pp. 23-44). Hillsdale, reducible to the other. NJ: Erlbaum. It was proposed that the representational Borsellino, A., & Poggio, T. (1973). Convolution and correlation algebras. Kybernetik, 13, 113-122. transformations, which occur because of the Bowerman, M. (1977). The acquisition of word meaning: episodic memory mechanisms in the holo- An investigation of some current concepts. In P. N. graphic model, CHARM, have several conse- Johnson-Laird & P. C. Wason (Eds.), Thinking (pp. quences. Among the most studied are some 239-253). Cambridge: Cambridge University Press. recall and recognition effects that have often Bradshaw, G. H., & Anderson, J. R. (1982). Elaborative been attributed to the depth of processing of encoding as an explanation of levels of processing. Journal of Verbal Learning and Verbal Behavior, 21, the items. These effects, as well as encoding- 165-174. specificity effects and both facilitative and Bransford, J. D., Franks, J. J., Morris, C. D., & Stein, inhibitory effects of elaboration, were ex- B. S. (1979). Some general constraints on learning and plained by the intrinsic similarities among memory research. In L. S. Cermak & F. I. M. Craik (Eds.), Levels of processing in human memory (pp. items and the highly interactive associative, 331-354). Hillsdale, NJ: Erlbaum. storage, and retrieval mechanisms in the Bransford, J. D., McCarrell, N. S., Franks, J. J., & CHARM model. Nitsch, K. E. (1977). Toward unexplaining memory. In R. E. Shaw & J. D. Bransford (Eds.), Perceiving, References acting, and knowing: Toward an ecological psychology Allwood, C. M., Wikstrom, T., & Reder, L. M. (1982). (pp. 431-466). Hillsdale, NJ: Erlbaum. Effects of presentation format on reading retention: Cavanagh, P. (1976). Holographic and trace strength Superiority of summaries in recall. Poetics, 11, 145- models of rehearsal effects in the item recognition test. 153. Memory & Cognition, 4, 186-199. Anderson, J. A. (1973). A theory of recognition of items Cermak, L. S., & Craik, F. I. M. (1979). Levels of from short memorized lists. Psychological Review, 80, processing in human memory. Hillsdale, NJ: Erlbaum. 417-438. Clark, E. (1973). What's in a word? On the child's LEVELS AND CHARM 37 acquisition of semantics in his first language. In T. E. Jacoby, L. L., & Witherspoon, D. (1982). Remembering Moore (Ed.), Cognitive development and the acquisition without awareness. Canadian Journal of Psychology, of language (pp. 65-110). New York: Academic Press. 36, 300-324. Craik, F. I. M. (1979). Levels of processing: overview Julesz, B., & Pennington, K. S. (1965). Equidistributed and closing comments. In L. S. Cermak & F. I. M. information mapping: An analogy to holograms and Craik (Eds.), Levels of processing and human memory memory. Journal of the Optical Society of America, (pp. 447-462). Hillsdale, NJ: Erlbaum. 55, 604. Craik, F. I. M., & Jacoby, L. L. (1979). Elaboration and Kinsbourne, M., & George, J. (1974). The mechanism distinctiveness in episodic memory. In L.-G. Nilsson of the word-frequency effect on recognition memory. (Ed.), Perspectives on memory research: Essays in Journal of Verbal Learning and Verbal Behavior, 13, honor of Uppsala University's 500th anniversary (pp. 63-69. 145-166). Hillsdale, NJ: Erlbaum. Klein, K., & Saltz, E. (1976). Specifying the mechanisms Craik, F. I. M., & Lockhart, R. S. (1972). Levels of in a levels-of-processing approach to memory. Journal processing: A framework for memory research. Journal of Experimental Psychology: Human Learning and of Verbal Learning and Verbal Behavior, 11, 671-684. Memory, 2, 671-679. Craik, F. I. M., &-Tulving, E. (1975). Depth of processing Kohonen, X, Oja, E., & Lehtio, P. (1981). Storage and and the retention of words in episodic memory. Journal processing of information in distributed associative of Experimental Psychology: General, 104, 268-294. memory systems. In G. E. Hinton & J. A. Anderson Eich, J. M. (1982). A composite holographic associative (Eds.), Parallel models of associative memory (pp. recall model. Psychological Review, 89, 627-661. 105-143). Hillsdale, NJ: Erlbaum. Estes, W. K. (1979). On the descriptive and explanatory Krinsky, R., & Nelson, T. O. (1981). Task difficulty and functions of theories of memory. In L.-G. Nilsson pupillary dilation during incidental learning. Journal (Ed.), Perspectives on memory research: Essays in of Experimental Psychology: Human Learning and honor of Uppsala University's 500th anniversary (pp. Memory, 7, 293-298. 35-60). Hillsdale, NJ: Erlbaum. Liepa, P. (1977). Models of content addressable distributed Eysenck, M. W. (1979). Depth, elaboration, and distinc- associative memory (CADAM). Unpublished manu- tiveneness. In L. S. Cermak & F. I. M. Craik (Eds.), script, University of Toronto. Levels of processing and human memory (pp. 89-118). Link, S. W. (1975). The relative judgment theory of Hillsdale, NJ: Erlbaum. choice reaction time. Journal of Mathematical Psy- Fisher, R. P. (1979). Retrieval operations in cued recall chology, 12, 114-135. and recognition. Memory & Cognition, 7, 224-231. Lockhart, R. S. (1979). Remembering events: Discussion Fisher, R. P., & Craik, F. I. M. (1977). Interaction of papers by Jacoby and Craik, Battig, and Nelson. In between encoding and retrieval operations in cued L. S. Cermak & F. I. M. Craik (Eds.), Levels of recall. Journal of Experimental Psychology: Human processing and human memory (pp. 77-88). Hillsdale, Learning and Memory, 3, 701-711. NJ: Erlbaum. Fisher, R. P., & Craik, F. I. M. (1980). The effects of Lockhart, R. S., Craik, F. I. M., & Jacoby, L. L. (1976). elaboration on recognition memory. Memory & Cog- Depth of processing, recognition, and recall: Some nition, 8, 400-404. aspects of a general memory system. In J. Brown Gabor, D. (1968). Holographic model of temporal recall. (Ed.), Recall and recognition (pp. 75-102). London: Nature, 217, 584. Wiley. Galton, F. (1879). Composite portraits, made by combin- Longuet-Higgins, H. C. (1968). Holographic model of ing those of many different persons into a single associative recall. Nature, 217, 104. resultant figure. Journal of the Anthropological Institute, Mandler, G. (1980). Recognizing: The judgment of pre- 8, 123-144. vious occurrence. Psychological Review, 87, 252-27). Gillund, G., & Shiffrin, R. M. (1984). A retrieval model Medin, D. L., & Schaffer, M. M. (1978). Context theory for both recognition and recall. Psychological Review, of classification learning. Psychological Review, 85, 91, 1-67. 207-238. Hintzman, D. L. (1983, June). "Schema abstraction" in Metcalfe, J., & Murdock, B. B., Jr. (1981). An encoding a multiple-trace memory model. Paper presented at and retrieval model of single-trial free recall. Journal the Conference on Priority of the Specific, Elora, of Verbal Learning and Verbal Behavior, 20, 161-189. Ontario, Canada. Moscovitch, M., & Craik, F. I. M. (1976). Depth of Hintzman, D. L., & Ludlum, G. (1980). Differential processing, retrieval cues, and uniqueness of encoding forgetting of prototypes and old instances: Simulation as factors in recall. Journal of Verbal Learning and by an exemplar-based classification model. Memory & Verbal Behavior, 15, 447-458. Cognition, 8, 378-382. Murdock, B. B., Jr. (1982). A theory for the storage and Jacoby, L. L., & Craik, F. I. M. (1979). Effects of retrieval of item and associative information. Psycho- elaboration of processing at encoding and retrieval: logical Review, 89, 609-626. Trace distinctiveness and recovery of initial context. Murdock, B. B., Jr. (1983). A distributed memory model In L. S. Cermak & F. I. M. Craik (Eds.), Levels of for serial-order information. Psychological Review, 90, processing and human memory (pp. 1-22). Hillsdale, 316-338. NJ: Erlbaum. Nelson, D. L. (1979). Remembering pictures and words: Jacoby, L. L., & Dallas, M. (1981). On the relationship Appearance, significance, and name. In L. S. Cermak between autobiographical memory and perceptual & F. I. M. Craik (Eds.), Levels of processing and learning. Journal of Experimental Psychology: General, human memory (pp. 45-76). Hillsdale, NJ: Erlbaum. 110, 306-340. Nelson, D. L., Brooks, D. H., & Wheeler, J. W, Jr. 38 JANET METCALFE EICH (1975). Sensory and meaning features in stimulus Ross, B. H. (1981). The more, the better?: Number of recognition and associative retrieval. Journal of Exper- decisions as a determinant of memorability. Memory imental Psychology: Human Learning and Memory, 1, & Cognition, 9, 23-33. 711-719. Schulman, A. I. (1974). Memory for words recently Nelson, D. L., Fbsselman, J., & Peebles, J. (1971). Words classified. Memory & Cognition, 2, 47-52. as phonological sequences. Journal of Experimental Smith, E. E., & Medin, D. L. (1981). Categories and Psychology, 87, 361-366. concepts. Cambridge, MA: Harvard University Press. Nelson, D. L., Wheeler, J. W, Jr., & Brooks, D. H. Sokal, R. R. (1977). Classification: Purposes, principles, (1976). Meaning and the elimination of sensory inter- progress, prospects. In P. N. Johnson-Laird, & P. C. ference. Journal of Experimental Psychology: Human Wason (Eds.), Thinking (pp. 185-198). Cambridge, Learning and Memory, 2, 95-102. England: Cambridge University Press. Nelson, K. (1974). Concept, word, and sentence: Inter- Thomson, D. M., & Tulving, E. (1970). Associative relations in acquisition and development. Psychological encoding and retrieval: Weak and strong cues. Journal Review, 81, 267-285. of Experimental Psychology, 86, 255-262. Nelson, T. O. (1977). Repetition and depth of processing. Treisman, A. (1979). The psychological reality of levels Journal of Verbal Learning and Verbal Behavior, 16, of processing. In L. S. Cermak & F. I. M. Craik (Eds.), 151-172. Levels of processing in human memory (pp. 301-330). Nisbett, R. E., & Wilson, T. D. (1977). Telling more than Hillsdale, NJ: Erlbaum. we can know: Verbal reports on mental processes. Tulving, E. (1979). Relation between encoding specificity Psychological Review, 84, 231-259. and levels of processing. In L. S. Cermak & F. I. M. Osgood, C. E. (1949). The similarity paradox in human Craik (Eds.), Levels of processing in human memory learning: A resolution. Psychological Review, 56, 132- (pp. 405-428). Hillsdale, NJ: Erlbaum. 143. Tulving, E. (1983a). Elements of episodic memory. Oxford, Posner, M. I., & Keele, S. W. (1970). Retention of England: Clarendon Press. abstract ideas. Journal of Experimental Psychology, Tulving, E. (1983b). Ecphoric processes in episodic mem- 83, 304-308. ory. Philosophical Transactions of the Royal Society of Raiijmakers, J. G. W., & Shiffrin, R. M. (1981). Search London, 302(R), 361-371. of associative memory. Psychological Review, 88, 93- Tulving, E., & Osier, S. (1968). Effectiveness of retrieval 134. cues in memory for words. Journal of Experimental Ratcliff, R. (1978). A theory of memory retrieval. Psy- Psychology, 77, 593-601. chological Review, 85, 59-108. Tulving, E., & Thomson, D. M. (1973). Encoding speci- Reder, L. M. (1982). Elaborations: When do they help ficity and retrieval processes in episodic memory. and when do they hurt. Text, 2, 211-224. Psychological Review, 80, 352-373. Reder, L. M., & Anderson, J. R. (1980a). A comparison Tversky, A. (1977). Features of similarity. Psychological of texts and their summaries: Memorial consequences. Review, 84, 327-352. Journal of Verbal Learning and Verbal Behavior, 19, Underwood, B. J. (1976). Recognition memory for pairs 121-134. of words as a function of associative context. Journal Reder, L. M., & Anderson, J. R. (1980b). A partial of Experimental Psychology: Human Learning and resolution of the paradox of interference: The role of Memory, 2, 404-412. integrating knowledge. Cognitive Psychology, 12, 447- van Heerden, P. J. (1963). A new optical method for 472. storing and retrieving information. Applied Optics, 2, Reder, L. M., & Anderson, J. R. (1982). Effects of spacing 387-392. and embellishment on memory for the main points of Wickens, D. D. (1972). Characteristics of word encoding. a text. Memory & Cognition, 10, 97-102. In A. W. Melton & E. Martin (Eds.), Coding processes Reed, S. K. (1973). Psychological processes in pattern in human memory (pp. 191-215). Washington, DC'. recognition. New \brk: Academic Press. Winston. Rosch, E. (1977). Classification of real-world objects: Wickens, T. D. (1982). Models for behavior: Stochastic Origins and representations in cognition. In P. N. processes in psychology. San Francisco: W. H. Freeman. Johnson-Laird & P. C. Wason (Eds.), Thinking (pp. Willshaw, D. (1981). Holography, associative memory, 212-222). Cambridge, England: Cambridge University and inductive generalization. In G. E. Hinton & J. A. Press. Anderson (Eds.), Parallel models of associative memory (pp. 83-104). Hillsdale, NJ: Erlbaum. Rosch, E., & Mervis, C. B. (1975). Family resemblances: Wittgenstein, L. (1958). Philosophical investigations (2nd Studies in the internal structure of categories. Cognitive ed.; G.E.M. Anscombe, Trans.). Oxford, England: Psychology, 7, 573-603. Blackwell. Rosenberg, S. (1968). Associative facilitation in the recall and recognition of connected discourse. Journal of Received September 12, 1983 Experimental Psychology, 78, 254-260. Revision received May 15, 1984 •